Histological coherent Raman imaging: a prognostic review

Abstract

Histopathology plays a central role in diagnosis of many diseases including solid cancers. Efforts are underway to transform this subjective art to an objective and quantitative science. Coherent Raman imaging (CRI), a label-free imaging modality with sub-cellular spatial resolution and molecule-specific contrast possesses characteristics which could support the qualitative-to-quantitative transition of histopathology. In this work we briefly survey major themes related to modernization of histopathology, review applications of CRI to histopathology and, finally, discuss potential roles for CRI in the transformation of histopathology that is already underway.

---

1 Introduction

Histopathology has its roots in sixteenth century biology,¹ and currently plays a key role in diagnosis of many cancers and other diseases. Nowadays, disease assignments from histology are generally based on the spatial arrangement of cell and tissue features that are highlighted using stains and sample preparation methods that were developed near the beginning of the 20^th century.² The current standard is subjective, consensus-based tissue analysis. This practice has held for many decades, however, it is known to suffer from several important sources of variability,^3–7 including differences in staining protocols and tissue analysis approaches among individuals and institutions. Further, while there are general agreements regarding sets of features that indicate a particular diagnosis,⁸ knowledge of and proficiency in identifying these vary significantly with level and emphasis of training among pathologists. Together, these factors lead to some degree of diagnostic imprecision. Given the central role of histology in influencing diagnosis and subsequent treatment regimens, it is not surprising that efforts are underway to increase objectivity, repeatability, and precision of histopathological analysis.

Over the past several decades, a number of approaches have been pursued to increase the amount of quantitative information extracted from histology samples, and thus support improved analytical performance of histopathology approaches. Immunohistochemistry⁹ (IHC), introduced in the early 1940s, enhances image information content through antibody-based contrast agents for diagnostically valuable targets, such as cell surface markers and structural proteins. This approach can yield information on the spatial distribution of these diagnostic targets and has proven very useful in some cases,¹⁰ but also has important limitations.^11–13 Other efforts have focused on quantifying image information and presenting it to physicians in the form of computer aided diagnostic (CAD) tools. A requirement for widespread use of CAD methods is that clinical tissue images be digitized. Although histopathology analysis is generally performed directly through the microscope, with no digital image ever acquired, digital recording of whole slide images (WSI) is becoming increasingly common, particularly for training purposes.¹⁴

Several label-free contrast modalities are being explored with regards to their utility in histopathology. Among these, vibrational spectroscopies such as Raman scattering and infrared absorption (IR) offer significant potential to increase the information content of histology images. Like IHC, these modalities provide chemically specific and diagnostically relevant information, but without the need of labeling. IR spectroscopy was applied as early as 1952 for qualitative characterization of normal and neoplastic tissue.¹⁵ Several IR and Raman histopathological studies in the 1990s through the 2000s utilized multivariate analysis methods to provide probabilities of disease within regions of interest.^16,17 Some studies also provided relative abundances of major tissue components having diagnostic value.¹⁸ While studies such as these showed impressive sensitivity and specificity, they often did not provide image information that could be easily integrated into a diagnostic decision. IR imaging was not of sufficient spatial resolution, and Raman spectral acquisition is sufficiently slow as to preclude imaging at the desired spatial resolution. These have been important limitations for a field where image interpretation is central.

The recent advent of coherent Raman microscopies¹⁹ have facilitated rapid label-free imaging, and bright infrared (IR) sources now available allow for IR imaging with higher spatial resolution.^20,21 When spatially resolved, vibrational spectroscopy can provide familiar and diagnostically important information such as cell arrangement, phenotype, cellular and subcellular morphologies, structural proteins, and intracellular lipids. Because they require little or no sample preparation, these modalities can provide this data in situ or with minimized latency between excision and availability of actionable information. Further, the intrinsically digital image format and potential for rich chemical contrast makes vibrational histopathology a strong candidate for strengthening CAD.

Here we review current applications of coherent Raman methods to histopathology, and consider its potential for meeting the needs of this field in the future. To better evaluate this potential, we first provide brief overviews of histopathology workflow, diagnostic criteria & uncertainty, and technology-related efforts underway to improve histopathology, including application of infrared and spontaneous Raman spectroscopy.

2 Histopathology

2.1 Histopathology workflow – a brief outline

Histopathology, the microscopic analysis of tissues for the purpose of identifying disease, may be performed on tissue samples pre-operatively, intra-operatively, or post-operatively. Pre-operative analysis is performed to help establish a detailed diagnosis and treatment regimen. Intra-operative and post-operative histopathology is performed to ensure complete resection of diseased tissue by confirming absence of disease in margins of extracted or in situ tissue. In cases where the benefit outweighs the risk, it is common practice to remove grossly healthy tissue surrounding the tumor to decrease the likelihood of positive margins. When excising a margin of nominally healthy tissue is contraindicated, such as for brain, the surgeon takes care not to remove more tissue than is justified. In such cases, intra-operative histopathology studies are sometimes performed on flash-frozen resected tissue to ensure that the entire tumor has been removed. The tissues are analyzed immediately, but an intra-operative histology analysis typically takes 20 to 30 minutes, and so prolongs the surgery.

Pre- and post-operative histopathology studies are typically done in a batch mode, and usually take a day or two to wend their way through the process, requiring approximately 13 hours of actual processing time. In this process, excised tissue is first grossly examined (by eye) for diagnostic information, and prepared for microscopy. Tissue samples are typically fixed and embedded within paraffin wax, and some of the tissue is sliced in 10 to 50 νm thick sections. These formalin-fixed paraffin-embedded (FFPE) sections are then stained (typically with hematoxylin and eosin [H&E]) and mounted for inspection. In a large hospital, a pathologist may inspect 500 histology slides in a day. Diseased regions are frequently small, so the clinician may spend a significant amount of time combing through normal tissue. In order to search more efficiently, pathologists will often identify landmark species such as blood cells and vessels, lymphocytes and ducts, which are frequently found in vicinity of cancerous lesions. Preliminary microscopic analysis and diagnosis are often carried out by a single pathologist. In difficult cases, a panel of experts may review the slides, and additional staining may be requested to identify diagnostically important species that H&E staining does not highlight. Ultimately, analysis and diagnosis are generally reviewed and signed-out by an attending physician.

2.2 Diagnostic criteria

The primary diagnostic criteria for most cancer types are related to intra- and intercellular organization.⁸ There are a number of additional criteria that are specific for various cancer types. These include cellular features such as number density of mitotic events, infiltration of immune cells, degree of cell differentiation within and surrounding the tumor, the degree of necrosis, abnormal nuclear shape or intracellular lipid levels, the appearance of over-developed mitochondria, and chromatin clumping.^8,22–24 Tissue structural features are also important for some cancer types, such as prostatic adenocarcinoma.²⁵ These features may include acinar organization, mineralized tissue, mucin or polysaccharide deposits, the presence or absence of structural proteins, such as laminin, elastin, or collagen. In the course of qualitative tissue classification, the pathologist must evaluate whether observed features fall in or out of the normal range. These judgments are subjective and thus susceptible to variability which is in a sense uncontrolled, because it cannot be quantified on a case-by-case basis.

2.3 Diagnostic uncertainty

While it is not currently feasible to assign diagnostic probabilities on a case-by-case basis, most studies address the issue of diagnostic uncertainty through analyzing outcomes in groups of patients. These studies can be classified into several types, and, while the results of these studies vary significantly, a few general trends emerge.

One type of study involves mandatory or voluntary review of a wide range of tissues by “peer” physicians (i.e., not necessarily subspecialty experts). In this class, Kronz et al.²⁶ and Raab et al.²⁷ report on inter-institution or institution-wide studies where large numbers of cases (6171 and 6162 respectively) covering many tumor types were reviewed. Kronz et al.²⁶ found 1.5% discordant diagnosis that resulted in a major modification of therapy or prognosis was observed. The majority of these cases involved a change between benign and malignant or a major change in tumor classification, and changes involving only a modification of tumor grade or stage were not included. That study found significant variability in diagnostic discordance with type of tissue. Serosal surfaces and the female reproductive tract tissues had 9.5% and 5.1% diagnostic discordance respectively, the highest found in the study. Raab et al.²⁷ estimated an error rate 6.7% from self-reported discrepancies upon second review at 72 institutions. In that study, only 1.1% of discordant diagnosis resulted in a major modification in therapy or prognosis.

Addressing these types of studies, Ho et al.²⁸ argue that statistics based on second review likely reflect an underrepresentation of errors since the reviewer often has knowledge of the original diagnosis and the sign-out pathologist. They argue that such a priori knowledge has led to biased review in similar situations.²⁹

In another class of study, previously diagnosed cases for a single tissue type are analyzed by a panel of sub-specialists. Higher discordant diagnosis rates are typically found in these types of studies. Lurkin et al.³⁰ report on a review of all sarcoma cases (366) in the Rhone-Alpes region of France over a 12 month period. They found that 19% of cases resulted in change of type or invalidation of diagnosis, and 27% of cases resulted in change of grade or subtype of diagnosed cancer. Bruner et al.³¹ found similar results for 500 neuropathology cases submitted to a specialist review committee. They found that 9% of reviews resulted in immediate significance for therapy or intervention, and 19% resulted in a change in type or grade of glioma. Similarly, specialist review of 602 prostate adenocarcinoma cases³² led to a change in the Gleason score by at least 1 point in 44% of cases, and patients’ risk category was increased in 11% of cases. Likewise, of 340 patients presenting for second opinions regarding breast cancer, 80% resulted in some diagnostic change, with 8% of reviews leading to altered surgical therapy.³³ In another study, 131 bladder carcinoma cases underwent secondary review, with 18% exhibiting significant discrepancies.³⁴ In a gynecological oncology³⁵ study of 295 referred patients, 5%, resulted in diagnostic changes that had major therapeutic or prognostic implications.

Arriving at definitive numbers for diagnostic uncertainty is outside the scope of this review. However, it does seem clear from the studies cited above that the uncertainty is significant for some cancer types. This widely recognized fact was reflected in a recent survey of pathologists,³⁶ which showed that physicians on average expect 10% diagnostic uncertainty from histopathology for cancer, and that lack of sub-specialty expertise is seen as the most important factor contributing to misdiagnoses. Based on these facts, one may be inclined to favor the sub-specialist review literature, and the slightly higher discordant diagnosis numbers found therein. The overall picture is that difficulty in diagnosis varies significantly with tissue and tumor type, and many human factors figure in precision and accuracy of diagnostics.^36,37

2.4 Impact of diagnostic uncertainty

Table 1 shows published figures for cancer rates, inter-observer diagnostic variability among pathologists, and cost associated with initial treatment. It appears that approximately 200 000 people each year in the U.S. are either incorrectly informed that they have cancer, and subsequently undergo invasive treatments, or are incorrectly told that they do not have cancer, and miss the opportunity for potentially life-saving treatments. One might assume that half of misdiagnosed cases are false positive and half are false negative. Under this naive assumption, one estimates the cost of unnecessary procedures is $2.5 B annually in the U.S. Of course, the cost of missed treatment is harder to estimate, but misdiagnoses are uniformly the largest reason for medical lawsuits, which cost $55.6 B annually, or 2.4% of all health spending.³⁸

Table 1 Major cancer types and diagnostic uncertainty

Type	New cases^39	Diagnostic uncertainty	Initial cost^40	Unnecessary treatment cost^a	Raman uncertainty
a These values are estimates assuming diagnostic uncertainty is evenly distributed between precision and accuracy.
Breast	250k	20%^7	$23k	$575 M	5%^18
Lung	225k	10%^3	$61k	$680 M	10%^41
Prostate	180k	30%^5	$20k	$540 M	5%^42
Colon & Rectal	135k	10%^6	$52k	$350 M	5%^42
Bladder	77k	40%^4	$21k	$320 M	7%^43
Melanoma	76k	25%^44	$5.5k	$52 M	5%^44

---

The magnitude and gravity of this problem is not lost on physicians and technologists, and the field of histopathology has been slowly evolving to provide better informed and precise diagnoses. Several innovations have been in process of adoption over the past decades.

2.5 Immunohistochemistry

Immunohistochemistry (IHC) typically uses fluorescently tagged antibodies to provide image contrast specific to proteins of diagnostic value. Most IHC tags are targeted to functional proteins, such as cell surface markers;⁴⁵ however, soluble⁴⁶ and structural^8,47–49 proteins are also targeted. The IHC approach facilitates mechanism-linked disease detection, and has enjoyed some important successes. For example, a protein (KIT), found to be mechanistically related to gastrointestinal stromal tumors (GISTs), is easily visualized through IHC labeling, and has found widespread use in diagnosing GISTs.¹⁰ While it now seems that KIT labeling is also prone to false positives in some cases,¹² the use of IHC for KIT detection has nevertheless simplified detection of GISTs, which was notoriously controversial with regard to classification, line(s) of differentiation, and prognostication.¹⁰ In spite of successes, IHC in general has some drawbacks including variability of labeling,¹¹ and incompatibility with some stains. Compatibility issues include background fluorescence from a stain that interferes with IHC detection, and stains chemically modifying the affinity antibodies used in IHC. H&E, for example, has both of these issues. Additionally, while IHC images can provide specific chemo-spatial information, they are also interpreted subjectively which can lead to variability in diagnostic outcomes.¹³

Improved diagnostic precision has been achieved by multiplexing classical and IHC stains, where the various contrast agents are imaged together.⁵⁰ Alternatively, cyclic IHC labeling^24,51–53 of samples facilitates registered imaging of many proteins. Such approaches can lead to improved specificity for cancer characterization, but can add considerable complexity and, in the case of serial IHC, require up to a day for each labeling cycle.⁵³

2.6 Whole slide imaging

Pathology slides are typically viewed and analyzed directly through a light microscope for diagnostic purposes. Image digitization is not a common practice for histopathology, but that is now beginning to change. Whole-slide imaging⁵⁴ (WSI), which entails digitization and storage of entire tissue slides, was introduced in the mid 1990s. Image acquisition speeds were initially too slow to be of practical use, but current WSI instruments have pixel acquisition rates on the order of 3 MHz, allowing them to generate bright-field images of a 1 cm² tissue sample at 500 nm resolution in about 2 minutes; roughly the time a pathologist might spend on a tissue sample of similar size. WSI is now used widely in teaching environments,¹⁴ and has been fully adopted in some hospital systems.⁵⁵ WSI is also useful for quality assurance, consultation, and telemedicine applications.^28,55

Fig. 1 shows an example of a prostate adenocarcinoma display from a pedagogical WSI application, allowing simultaneous view of an image thumbnail, and a detailed sub-image section.¹⁴ This sort of data presentation has obvious pedagogical and practical benefits, but large, high-resolution images such as this may occupy ≈1 Gb of disk space. Thus, the work of a single pathologist may occupy 0.5 Tb day⁻¹, and storage space requirements are often cited as a potential barrier to widespread acceptance of WSI.⁵⁶ Another barrier is slightly reduced image quality compared to direct microscopic observation.⁵⁵ There are further concerns about variation in sample preparation and staining practices, and that image capture and coloration can vary over an important range among WSI instruments. Consideration of these issues is currently underway at the U.S. Food and Drug Administration.⁵⁷

Fig. 1 WSI image of prostate adenocarcinoma tissue from set of Genitourinary Block training in the preclinical curriculum at the University of Arizona College of Medicine, in Tucson, AZ.¹⁴

2.7 Computer aided diagnosis (CAD)

CAD approaches support improved diagnostic precision by providing image-derived, diagnostically relevant information to pathologist. Efforts towards CAD began in the late 1990s with detection of malignant masses in radiological images of breast tissue,⁵⁸ and CAD is now routinely used in radiology for identification of lesions. Further, it seems that the prevalence and sophistication of CAD is likely to grow. In 2015, IBM acquired Merge Healthcare, for its 30 billion radiological and histological images, with the intent of using the image data to train Watson, their physician assistant software.⁵⁹

While radiological CAD is now well-integrated into the diagnostic process, histological CAD is not used routinely. This dichotomy may be due in part to the fact that the former simply provides evidence for the presence of tumors, while the latter must address the more complex questions of tumor type and grade, and deals with images of higher complexity. Numerous algorithms have been proposed for diagnosis of specific cancers,^60–62 but there has not been significant clinical adoption. Histological CAD, however, seems to be emerging from the research-only phase, as a number of companies are becoming involved in developing these tools.⁶³

Histological CAD systems may exploit a large range of image features to derive clinically significant information,⁶⁰ including object size and shape, intensity, color, texture, and inter-object organization. Approaches for extracting these inputs from images have been reviewed recently.^64,65 The image processing associate with CAD will generally include some or all of the following steps: (1) preprocessing, such as color and intensity normalization and de-noising. (2) Feature identification using thresholding or segmentation approaches. (3) Quantification of relevant feature characteristics, such as shape, size, spatial arrangement, texture, or color. (4) Reduction of data dimensionality using methods such as principle component analysis. (5) Classification based on the reduced-dimension dataset. Steps 2 through 5 (and sometimes step 1) are collectively referred to as “machine learning”. We briefly discuss these steps below.

2.7.1 Preprocessing. Image analysis and quantification ultimately depend on spatial variation in color and intensity. It is therefore crucial that variations in tissue properties of interest give rise to the same range of image color and intensity variations within a sample, or between samples. However, in practice, it is often difficult to achieve sufficient image uniformity. Many factors such as inconsistency in staining or storage conditions, illumination brightness and uniformity or sample tilt can induce significant variation in response of the captured image to actual changes in tissue properties.

Use of blank images⁶⁶ or reference materials can reduce inter-image variability, but this is often insufficient. In order to minimize these variations, preprocessing steps such as color and intensity normalization are often employed. Color convolution⁶⁷ quantifies staining densities of a small number of contrast agents based on the fact that light absorptivity of individual stains is linear in log space. This approach continues to be developed, and can be used to map image color and intensities from many samples onto common range,⁶⁸ assuming each image contains similar ranges of features. However, color and intensity normalization approaches can lead to significant artifacts when based on assumptions that the proportion of pixels or range of stain intensities from each image in a set should be identical.⁶⁹ In cases where a given stain vector may not be appropriate for all samples, individual vectors can also be normalized to a single image.⁶⁹

2.7.2 Image segmentation. Segmentation is the process of grouping like pixels into super-pixels, and generating boundaries around regions or objects with similar properties in the image. There are many segmentation approaches. The simplest of these is thresholding, where pixels in an image are converted to 0 or 1 depending on whether they meet a threshold criterion. The threshold may be set by consideration of pixel values in the overall image⁷⁰ or of local regions if the image is expected to have trivial nonuniformities.⁷¹ Watershed⁷² is another common segmentation approach where the boundaries of enclosed object shapes are defined by local maxima in pixel values. K-means clustering is another approach which finds, for some number (k) of clusters, the cluster location and boundaries that minimize a merit function, designed, for example, to group pixels of similar intensity. Fuzzy clustering^73,74 and Gaussian mixture models⁷⁴ allow for the possibility of overlapping objects, assigning each pixel to one or more contiguous objects that are defined in such a way to minimize variation of a particular property within the boundaries of the object. The optimal configurations of these clusters are often found using expectation maximization approaches.^74,75

Another approach uses active contours, which are deformable splines that encompass image segments. They are constructed to minimize some energy function that may have contributions dictating the stiffness of the spline, the placement of the spline on the image (e.g., lowest energy at a minimum image gradient location), and a term that encapsulates a priori knowledge of image segment shape or location.⁶⁵ These splines can be parameterized (SNAKES)⁷⁶ or unparameterized (level set methods).⁷⁷ Each approach has strengths and weaknesses, with SNAKES being more robust to noise, and level set methods having greater flexibility to accommodate variable topology. In cases where salient aspects of the image segment shape is known, Bayesian analysis of template comparison⁷⁸ has been successful, and found robust to image variation. Each of these segmenting approaches may have trouble with overlapping features of interest, and active contours are prone to inclusion of unwanted background objects.⁶⁴ Also, Bayesian methods are, in general, computationally expensive.

Of course, pixel-level and active contour approaches are not mutually exclusive. It is possible to leverage the strengths of each in a single algorithmic approach.⁷⁹

2.7.3 Feature selection. Image segmentation prepares one to select features that may be used for quantification or in predictive model construction. This step frequently reduces complexity of the data to be considered and may be performed with or without user supervision. Supervised methods require the user to specify properties of features to be found, whereas unsupervised methods discover these properties autonomously from the data. The latter class includes principle component analysis (PCA),⁸⁰ which finds a minimum set of orthogonal, linear combinations of the original data dimensions that retain the majority of signal variation. Exploratory factor analysis (EFA)⁸¹ is a similar approach, but differs in that it allows for a component of random error in apparent associations between features and observed variables. Independent component analysis (ICA)⁸² is also similar to PCA, except that ICA attempts to unmix signals by using the non-Gaussian character of individual signal elements to separate them into components that are as statistically independent from one another as possible.

Linear discriminant analysis (LDA) is a commonly used supervised feature selection approach that assumes independent variables are normally distributed. It is similar to PCA, except that the feature types are specified by the user. A comparison of PCA and LDA can be found in Martínez et al.,⁸³ who prefer PCA when the training dataset is small, other factors being equal.

Similar to methods mentioned above, manifold learning (ML) approaches project an M dimensional dataset to a space with fewer (N) dimensions, where preferably, N ≪ M. Both supervised and unsupervised ML approaches exist, and these differ from those above in that there is no assumption of linearity in the relationship between the features of interest and the original data.⁶⁴

2.7.4 Feature quantification. Once diagnostically important image features are identified, their properties may be quantified. Spatial arrangement of cells is almost always important, and the Feldman group^78,84 provide an example of how this can be characterized. They start with Voroni tesselation, in which the image is partitioned into convex polygons such that each polygon contains exactly one nucleus and all points within the polygon are closer to the center of their nucleus than to that of any other. From this starting point, they use Delaunay triangulation to draw lines between nuclei that shared a Voroni boundary, and then choose a set of lines that form a continuous structure of minimum lenght – a minimum spanning tree. From characterization of the tree and a dozen other similar metrics, were able to discriminate Gleason 3, 4, and normal regions for prostate,⁸⁴ and evaluate Bloom-Richards criteria for breast cancer.⁷⁸ Automated analysis and classification compared favorably to a manual approach, yielding better results in every categorization except cancer vs. non-cancer in breast tissue.

2.7.5 Classification. Once image features have been quantified, the associated metrics may be used to classify tissues regarding their disease state. One approach is to use decision trees, which are a series of questions, the answers to which lead directly to a classification. These approaches have a number of attractions; they are similar to the approach a human might use to classify a tissue and are transparent. Another is that they have low systematic bias.⁸⁵ A potential drawback is that they tend to have high variance, but this can be reduced using approaches such as random forests,^85,86 where many decision trees (the forest) are grown on varying subsets of the data, and decision nodes of the final tree are based on statistical sampling of the nodes in the forest.

Deep learning is another route to classification. These approaches are an evolution of neural networks in which the data vectors are iteratively transformed using nonlinear functions in such a way as to emphasize image aspects having predictive power. Using these approaches, segmentation, feature selection, and classification are all done at once, and all without user intervention. However, the entire process is opaque to the user in that there is no way to know what features the model is using for classification. Thus, there is thus no way to know when the model is likely to fail. In fact, only to the extent that the dataset on which the deep learning model was trained is comprehensive, can the characterization be reliable. By way of apparent counterexample, however, Cirean et al. were able to use a relatively small sampling of mitotic events to create a highly effective deep learning model for detecting mitoses in H&E images.⁸⁷ Nonetheless, building a comprehensive dataset for deep-learning image recognition seems to be the intent of IBM in acquiring a company with access to 30 billion radiological and histology images.⁵⁹

2.7.6 Roles for CAD in histopathology. CAD assists the pathologist in one or more of the image analysis steps leading to diagnosis. The assistance could range from simply providing quantitative image metrics, to contextual reminders of disease mimics to apparent probabilities of various diagnoses.

Even the most basic level of CAD assistance – that of providing image feature identification, such as of nuclei or regions of potential metastases⁸⁶ may be quite valuable. In many cases, a pathologist may search for regions of interest (ROIs) that cover only a small fraction of the sample. Identifying such regions up front, or even identifying landmark species that typically signify such regions could provide a significant time savings.⁵³

Feature quantification might be considered the next level of sophistication where CAD approaches could be of high value. We find evidence for this by considering the work of Fuchs and Buhmann,⁸⁸ who found 10% disagreement on location and size of nuclei in a renal cell carcinoma sample between two sub-specialists, and inter- and intrareader disagreement of 42% and 21% respectively for counts of normal and atypical nuclei among five pathologists who were not sub-specialists. This level of variability is comparable to the overall diagnostic uncertainty discussed in Section 2.3, and suggests that correct identification of relevant features may play a significant role in diagnostic uncertainty for this and other cancers. Image analysis approaches discussed in the previous sections can provide quantitative characterization of key features, such as cell density⁶¹ or mitotic event counts⁸⁷ and nuclear shape metrics^78,84 at precision levels similar to that of sub-specialist analysis in the Fuchs and Buhmann work.⁸⁸

While CAD work is largely done using H&E,⁸⁷ improved approaches to image contrast would likely lead to increased reliability of CAD outputs.⁶³ Naturally, an expanded contrast pallet serves to put image-derived metrics on a more solid footing.⁵³ Even the simplest pixel-level thresholding methods can be quite reliable when pixel intensity values are related to specific chemical or functional information. For example, pixel-level analysis of vibrational spectra resulted in accuracies of 94%–100% for classification of ten disparate histologic classes,⁸⁹ 98% for positive prostate nuclei from IHC images with contrast for androgen receptor protein,⁹⁰ and 95% for micro-metastasis from cytokeratin-stained lymph node sections.⁹¹

2.8 Emerging contrast modes

While use of classical stains and IHC approaches is well-established in histopathology practice, interactions of light with unlabeled tissue can also be used as contrast to provide important diagnostic information. Elastic interactions between light and tissue are generally characterized through static light scattering or quantitative phase imaging (QPI). Inelastic interactions are characterized through light absorption, autofluorescence, and vibrational spectroscopy. If one could rely exclusively on these label-free contrast modalities, there could be a significant reduction in sample-to-sample contrast variability in addition to reduced costs and handling time of pathology slides. It is unlikely that label-free contrasts alone could, in every case, provide sufficient information for diagnosis. However, these label-free methods do provide a surprising amount of diagnostic information – perhaps sufficient for many or most cases.

Refractive index maps and light scattering patterns from tissues both provide information useful for diagnosing aspects of some cancers. In light scattering experiments, light is refracted, or scattered, as passes through materials with discontinuities inrefractive index. Optical coherence tomography is now well developed as a clinical imaging tool, and can be used to determine spatial location of scattering sites.⁹² Additionally, polarization, wavelength, and scattering angle may be measured directly in an imaging mode to deduce relevant tissue properties. Tumor growth modes produce tissues that scatter light strongly⁹³ and anisotropically,⁹⁴ and it appears that light scattering may be sufficiently sensitive for very early tumor detection of some epithelial cancers.⁹⁴ Elastic interactions of light and matter may also be characterized by directly mapping out the refractive index of a sample through quantitative phase imaging.⁹⁵ QPI reports on refractive index variation from the nano-scale to the micro-scale, and as with qualitative phase imaging, it makes boundaries between tissue regions and cellular compartments readily visible. QPI can also discriminate some tissue and cell types,^93,96 locate micro-calcifications,⁹³ and even indicate degree of malignant transformation⁹⁷ based on values of refractive index. Light scattering pattern maps and phase imaging maps are essentially interchangeable as the former can be calculated from the latter and vice versa, provided the light scatting is spectrally resolved.^98,99

Light absorption¹⁰⁰ and intrinsic fluorescence¹⁰¹ from ultraviolet¹⁰² or visible and infrared light¹⁰³ can be used to characterize tissues. When specific excitation and emission wavelengths are used, signals deriving primarily from species such as nicotinamide adenine dinucleotide (NAD) and flavin adenine dinucleotide (FAD),¹⁰⁰ elastin and collagen,¹⁰⁴ tryptophan,¹⁰⁵ and possibly porphryns¹⁰³ can be isolated. Accumulation or deficit of these species is not specific to tumors, but does seem to be a sensitive indicator for the presence of neoplastic masses. In fact, autofluorescence has been used as method to rapidly locate candidate tissue regions for a slower, but more specific mode of characterization, vibrational spectroscopy.¹⁰⁶

3 Spectroscopic histopathology

Vibrational spectroscopy characterizes molecular vibrations intrinsic to species of interest in a specimen. This rich source of chemical information has been applied to histopathology, albeit primarily in a research mode. From spectroscopic and spectral imaging studies it appears that much of the information sought through staining, and perhaps more, is available without labeling from techniques such as Raman or IR spectroscopy.

It is common to describe vibrations of molecules with N constituent atoms in terms of oscillatory motions along vectors with up to 3N-6 dimensions called “normal modes” (Q). These modes constitute collective intramolecular motions that preserve the overall symmetry of the molecule. The modes have discrete allowed states with energies approximately equal to (ν + 1/2)ħω_m, where ν are integer values, and ω_m is the ground vibration frequency. The ground state vibrational frequencies and energies vary depending on the strengths of bonds and masses of atoms involved in the normal mode motion. Transitions with Δν = ±1 are most often measured, and those with energies in the range (500 to 3100) cm⁻¹ are typically of interest in biological systems. Signatures in the higher energy range (2800 to 3100) cm⁻¹ arise from transitions between states of modes involving symmetric or asymmetric stretching of C–H bonds. The range (1800 to 2800) cm⁻¹ is referred to as the quiescent region because transitions of natural biological molecules typically do not appear in this range. The greatest variety of vibrational transitions in biological molecules occur in the fingerprint range (500 to 1800) cm⁻¹.

Transitions between these states are usually detected through absorption, emission, or inelastic scattering of light. The probability for photon absorption or emission is significant when ∂μ/∂Q ≠ 0 at Q = 0, where μ is a mode's permanent dipole moment. By contrast, inelastic scattering is important when ∂α/∂Q ≠ 0 at Q = 0, where α is the molecular polarizability. Infrared and Raman spectroscopies are respectively based on absorption and inelastic scattering.

Fig. 2, panels a and b show IR and Raman spectra for major classes of tissue constituents. Since these have distinct signatures in both spectroscopies, they can be identified by either method. On the other hand, IR absorption linesare slightly broader. This is of consequence in the fingerprint region, where spectral lines are quite congested. Panels c and d of Fig. 2 show fingerprint regions of IR and Raman spectra from cells, illustrating that the width of the individual IR peaks makes it difficult to separate them into more than just a few distinct contributions without peak fitting. On the other hand, Raman peaks are more narrow, so that spectra from similar systems contain a larger number of easily resolvable peaks, providing a much higher level of chemical specificity. The differences can be quantified in terms of information content, or entropy, S_i = ln(Ω_i), where Ω_i is the maximum number of distinct spectral states one can discriminate. Here, Ω_i = SNR^M where M is the number of spectral peaks and SNR is the signal to noise ratio in each peak. We can estimate this value by noting approximately 15 peaks in the IR spectrum, 5 peaks in the Raman CH region, and 45 peaks in the Raman fingerprint region. Assuming SNR = 15 for IR peaks and for CH stretch Raman peaks, and SNR = 3 for Raman fingerprint, we obtain Ω_i = 10¹⁷, 10⁶, and 10²¹ for IR, Raman CH, and Raman fingerprint spectra respectively (10²⁷ for the entire Raman spectrum). Partly for this reason, Raman is found to have better specificity for classifying cancer.¹¹⁰

Fig. 2 a & b: Respectively, IR spectra and Raman spectra of major cell components.¹⁰⁷ From top to bottom on both panels, spectra are from albumin, globulins, collagen, glycogen, DNA, RNA, and lipid. c: IR fingerprint spectrum of whole cells,¹⁰⁸ with ≈12 resolvable peaks. d: Raman fingerprint spectrum of whole cells,¹⁰⁹ with ≈45 resolvable peaks.

Given their capacity for encoding chemical information, it is not surprising that these spectroscopies have been applied to many general aspects of disease diagnoses.^111,112 They have also been used for guidance in tissue sampling,¹¹³ augmented bases for stratified diagnosis,¹¹⁴ and intraoperative guidance during tissue-conserving resection.^115,116 Of particular present interest is the considerable body of literature showing that vibrational spectroscopy provides diagnostically significant information for cancer diagnosis in conjunction with histopathology. There have been many excellent reviews on this topic in recent years,^{107,111,117–121} so here we will only summarize relevant aspects.

Histology-related studies using vibrational spectroscopy began to appear in the 1960s¹²² for IR, and in the 1980s for Raman,¹²³ and to date, there have been thousands of publications. However, it appears that only in the past decade or so that IR and Raman technology have been sufficiently developed to produce meaningful studies.¹⁰⁷ Recent work provides rather convincing evidence that vibrational spectroscopy can detect diagnostically important changes in tissues, routinely yielding objective sensitivity and specificity figures generally in the 85% to 95% range for detection and classification of cancers, even when spectra on which discrimination is based are usually taken at just a few spots from normal and diseased portions of tissue.

In one example, Pence et al.¹²⁰ discuss 26 large studies of epithelial cancers (>50 patients), among which, the average sensitivity and specificity obtained using Raman spectroscopy were 89 ± 7% and 88 ± 8% respectively. Such precision generally exceeds that possible by subjective morphological analysis. Similar examples are listed in the final column of Table 1, indicating a reduction in diagnostic uncertainty by more than a factor of two in most cases. Table 1 also provides prevalence, histology-based diagnostic uncertainty and initial treatment costs for the six most common cancers in the US. Assuming naively that diagnostic uncertainty is evenly distributed between precision and accuracy, we estimate that using Raman spectroscopy to supplement histology could save >$2B per year in unneeded treatments, and reduce the burden on patients and families associated with missed opportunities for earlier treatment.

3.1 Vibrationally detected cancer markers

The impressive diagnostic value of vibrational spectroscopy accrues from its inherent chemical specificity. It can reveal information that is otherwise available only by methods such as IHC and polymerase chain reaction (PCR) or gene arrays,¹⁰⁷ leading to more precise tissue segmentation,⁸⁹ and providing new prognostic information.¹³³

Table 2 gives a partial list of cancer biomarkers that have been used in Raman studies to classify neoplastic and cancerous tissues, and lists spectral peaks used to identify those markers. Most markers have several peaks, facilitating increased robustness in detection, to the extent that noise in separate peaks is uncorrelated, but even correlated noise can be suppressed if peak ratios can be used. The fact that most of the peaks used in these Raman studies are found in the fingerprint region is due to the high spectral density and low peak widths found there, and underscores the importance of this spectral range for chemical specificity.

Table 2 Major cancer markers detected through vibrational spectroscopy

Marker	Relevant cancers	Spectral lines [cm^−1]
ν-stretch, δ-scissoring, ρ-rocking, ω-wag, τ-twist.
Metabolic markers
DNA (relative abundance)	Breast^18 larynx^124	668, 678, 728, 750, 785νO–P–O 825νO–P–O, 1336, 1488, 1580
Lipid (structure and relative abundance)	Breast^18 colon^42,125 bladder^43 skin^44,126	1309τCH2 1445δCH2 1654νC–C 2850νCH2,sym 2880νCH2,asym 2920νCH3,sym 2960νCH3,asym
Cholesterol ester	Breast^18 brain^127	430, 702, 1302, 1442, 1740νC=O
Glycogen	Bladder^43 prostate^42 brain^127	472, 481, 846, 932
Creatinine	Breast^128	600νN–CH3 678νC=O 685τC–S 692δCO2 840δN–CH2 903νC–C–N
Structural molecules, etc.
Protein structure	Skin^44 melanoma^44	1660νC–O (amide I) 2850νCH2,sym 2880νCH2,asym 2920νCH3,sym 2960νCH3,asym
Keratin	Epithelial cancers^129	623νCS, 643νCS, 850δCCHaromatic, 885ρCH2 933, 1002νC–C (ring breathing), 1031 (phenylalanine), 1200 1350 (amide III) 1450τCH 1650νC=O (amide I), 2940νCH3
Collagen	Breast^18 prostate^42 colon^42 lung^41,130 skin^131	855, 938, 1004νC–C1260νCN, δNH1314ωCH3CH21445δCH21660νC–O 2940
Elastin	Lung^41	1442δCH2 1660νC–O
Carotenoids	Bladder^43 lung^41 brain^127	1159, 1523
Polysaccharides	Brain^132	856
Others
Phospholipids	Lung^41 brain^127	719 1442δCH2
CaC2O4, Ca5(PO4)3	Breast^18 brain^132	912νC–C 960νP=O 1477νC–O
Phosphatidylcholine	Adenocarcenoma^107	719, 1666νC–O
Trypotphan	Lung^41 prostate^42	650τC–C 1260νCN, δNH
Proline	Prostate^42	939νC–C
Tyrosine	Prostate^42	1176νC–H 1217νC–C6H5

---

Many of the markers listed in Table 2 are commonly visualized in pathology through IHC or classical stains, but, of course those methods require labeling. Not only is vibrational contrast label-free, it can also be semi-quantitative when peak amplitude ratios are used. Accordingly, peak pairs that provide information on protein-to-DNA or protein-to-lipid ratios, or on degree of unsaturation in fatty acids frequently yield highly reliable metrics of disease type and grade.^{41,124,132,134–137}

Macromolecular structure and material phase information are also available from vibrational spectroscopy, and can be useful in cancer detection and diagnosis. For example, degree of calcification is a known marker for many cancers.⁸ Using classical stains, it is difficult to distinguish between types of calcification, but such discrimination turns out to be important. Haka et al.¹⁸ showed that Raman spectroscopy can be used to easily distinguish CaC₂O₄ from Ca₅(PO₄)₃, and that their relative abundances correlate with malignancies in breast cancer. Raman spectroscopy also contains structural and hydrogen bonding information for lipid, protein, and nucleic acids, which has diagnostic value for many cancers.^{18,44,126,128}

Spectral changes between healthy and diseased tissue can be striking,^41,138 or subtle,¹⁰⁷ but generally appear in the context of highly complex spectra from the tissue. Thus, while it is sometimes possible to determine identity of diagnostically relevant species though differences between spectra from normal and diseased tissue,¹²⁷ this is sometimes not possible even when the primary marker species is known.¹²⁹ For this reason researchers often derive diagnostic information from tissue spectra using spectral pattern recognition approaches.

3.1.1 Machine learning on spectral features. Many of the machine learning methods used for image feature selection (see Section 2.7.3) are also used for spectral feature selection and classification. Unsupervised feature selection methods such as principle component analysis (PCA, linear transform), singular value decomposition (SVD, linear transform), and kernel PCA (KPCA, nonlinear transform) generate orthonormal basis vectors containing correlated spectral features.^139,140 These basis functions may contain positive and negative spectral elements, so their interpretation is not always clear. However, they are useful for denoising or dimensionality reduction, in which only a subset of basis vectors are retained and spectral information of interest is reconstituted through linear combinations of the retained vectors.¹⁴¹ Other unsupervised feature selection methods, such as nonnegative matrix factorization^140,142 (NMF; also ‘positive matrix factorization’), positivity constrained multivariate curve resolution^{140,143–145} (MCR) and vertex component analysis^139,146 (VCA), are designed to retrieve physically meaningful, positive-valued spectra from which abundance maps can be directly generated. MCR and NMF techniques return components of spectra that carry discriminating information, and may or may not return realistic spectra of molecular constituents, depending on their implementation. VCA assumes that there are ‘pure pixels’ for each molecular entity (i.e., the unmixed spectrum of each pure chemical component is represented at least once among all the spectral pixels). Methods that are explicitly designed to return spectra corresponding to real molecules are termed ‘endmember extraction’ or ‘spectral unmixing’.

These unsupervised feature extraction methods frequently serve as the input to unsupervised and supervised classification methods. Unsupervised methods, such as K-means clustering^139,146 (KMC), divisive correlation cluster analysis (DCCA), and agglomerative hierarchical cluster analysis (AHCA), link together related spectra based on spectral proximity metrics.¹³⁹ Supervised methods, such as linear discriminant analysis¹¹¹ (LDA), partial least squares^147,148 (PLS), support vector machines^149,150 (SVM), and random forest classifiers,^146,151,152 rely on training data to generate a classification model. These models may be conservative or relaxed according to clinical need. For applications such as tissue-conserving resection,¹¹⁵ one may choose a conservative model to minimize false positives. Alternatively, if the aim is to reduce the burden on pathologists, one could design a model that may include false positives, but automatically classifies the most clear-cut cases, leaving the rest for inspection.¹⁵³ Training these models may be arduous, requiring many datasets and extensive validation; but post-training classification may be rapid.^146,148 There are, additionally, ‘semi-supervised’, methods that aim to incorporate known and unknown information into classification problems, such as hybrid linear analysis¹⁵⁴ (HLA) semi-supervised PCA-linear discriminant analysis¹⁵⁵ (PCA-LDA).

Deep learning¹⁵⁶ (DL) and artificial neural networks^157,158 (ANN) constitute an important family of methods that have received significant attention recently. These methods rely on layers of interconnected decision units or ‘neurons’; typically 1–2 layers for ANN approaches, and potentially many layers for DL approaches. Whether employed as supervised or unsupervised learning, the weights and the interconnection strengths of the neurons are iteratively modified to optimize a merit function. One of the primary strengths of ANNs and DL is that nonlinear and extremely complex interconnections can be created within the network that may not be possible with traditional linear or nonlinear methods such as those mentioned above. A significant challenge to ANNs and DL, as with all supervised techniques, is the requirement that the training set be comprehensive in that it is representative of all future data. Another challenge is that these approaches are ‘black box’ in that it is not currently clear how to understand what data features are found to be important, and how they are used in the evolved ANN/DL neural connection architecture. Interpretation of ANNs and DL is an active area of research.¹⁵⁹

3.1.2 Spatio-spectral analysis. The diagnostically relevant information contained in vibrational spectroscopy from tissue samples is useful for histology to the extent that it can be spatially resolved on relevant length scales. Raman and IR spectroscopies are both conducive to spatially resolved acquisition, and both have been used in an imaging modality with tissues.

Thanks to recent commercial availability of quantum cascade lasers, approaches²⁰ for IR imaging at spatial resolution as high as 1 μm are now feasible, making it possible to generate gross tissue maps based on IR absorption. Since the chemical specificity of IR spectroscopy is sufficient to easily discriminate among tissue regions such as epithelium and stroma,¹⁶⁰ and between benign and malignant tissue,⁸⁹ IR maps can be merged with H&E images to indicate tissue regions of interest,¹³³ as shown schematically in Fig. 3a. Morphological features extracted directly from such maps have been used to classify tissue samples as cancer or non-cancer, with high accuracy.¹⁶¹ This automated ability could provide considerable value, as mentioned in Section 2.7.6.

Fig. 3 IR and Raman histopathology: (a) Infrared absorption spectra contain sufficient information for classification of tissue, and provide value through indicating regions of interest.¹⁶¹ (b) A: H&E stained section of lymph node germinal and mantle zone, and B: a sub-image with contrast from Raman spectroscopy showing activated lymphocytes in yellow.¹⁰⁷ (c) Spatially multiplexed Raman spectroscopy with 30 ms effective spectral acquisition time using 500 mW of 532 nm light. The cytochrome-c distribution is indicated in yellow.¹⁶⁴

From a spatial resolution perspective, Raman imaging has an important advantage over IR in that it uses visible or near infrared wavelengths, so can reach spatial resolution 2× to 3× higher than with IR, making cellular and sub-cellular imaging possible. The combination of high resolution and rich fingerprint spectra allows for significant information in addition to providing classification of tissue regions. For example, Raman imaging can provide label-free identification of cell type,¹⁶² distribution^139,163 and content of organelles.¹⁶⁴ Important functional information is also available, such as DNA to protein ratio of nuclei,¹⁶⁵ the cell cycle state,^163,166,167 and whether immune cells such as leukocytes are activated.^107,168 Examples of cellular function and organelle distribution maps generated with Raman spectroscopy are shown in Fig. 3b and c.

Although IR and spontaneous Raman imaging can provide useful diagnostic information, both have practical drawbacks. For example, conventional glass slides can be problematic for both methods as they emit fluorescent light when excited with visible and near IR light, and absorb IR light at wavelengths of 3.5 μm and longer, which constitute the fingerprint region for IR. Consequently, both approaches often require special substrates – quartz or CaF substrates for Raman studies, and CaF or BaF₂ substrates are used in IR. On the other hand, glass slides are transmissive to IR wavelengths in the range λ = (2.8 to 3.5) μm, allowing only the CH stretch spectral region to be used.¹⁶⁰ Another consideration is that the intrinsically low Raman scattering levels can be easily masked by fluorescence; thus, in addition to use of specialty substrates, it is often necessary to photobleach samples before acquiring Raman spectra.

The benefits of spectral image acquisition should be considered in context of required imaging time. These considerations differ significantly for IR and Raman spectral imaging. IR absorption cross sections for fingerprint and CH stretch modes are typically on the order of 10⁻¹⁸ cm², so most of the light will be absorbed for a high density of absorber with (5 to 10) μm path length. Having such a large effect on transmitted light, signal can be acquired very quickly. Yeh et al.²¹ describe work wherein 128 × 128 pixel wide field IR absorption images were acquired serially on a fixed region of the tissue while the laser wavelength was scanned. In total, 282 frames were acquired, yielding IR absorption images over a 1128 cm⁻¹ range in the fingerprint region, with an effective pixel rate of 1.6 MHz, and spectral acquisition rate of 3 kHz. In cases where fewer spectral points are required, some imaging speed increase can be obtained by using only several discrete frequencies. For example, Tiwari et al.¹⁶⁹ found that they could reliably identify diseased cardiovascular tissue regions with on-the-order of 10 discrete spectral points. Allowing for imaging of ∼10 discrete contrast frequencies and a 5 μm pixel size, tissue could be mapped at 0.5 cm² min⁻¹, similar to WSI instruments currently in use. Further, if the discrete frequencies are exclusively in the CH stretch region, glass slides could be used, and there would be minimal impact on the histology work flow, as IR imaging can be performed on H&E stained slides.¹⁶⁰

In contrast to IR absorption, the Raman signal is quite small, and imaging with this contrast mechanism is proportionally slow. The differential scattering cross-section of Raman-scattered light collected of over solid angle (Ω) is given by dσ/dΩ ∝ |∂α/∂Q|², and is on the order of 10⁻³⁰ cm² sr⁻¹ for most modes. Accordingly, spectral acquisition rates typically range from 5 to 0.01 Hz.^129,170 Using spontaneous Raman scattering, one cannot excite only selected vibrational frequencies, so there is no benefit to discrete spectral imaging as in IR. On the other hand, some increase in acquisition speed can be achieved through spatial multiplexing. Fig. 3c displays a spontaneous Raman image acquired at a 90 Hz effective spectral acquisition rate¹⁶⁴ through line-focused (rather than spot-focused) excitation light. Because practical detector arrays are presently limited to 2 dimensions, and one of those is devoted to spectral variations, spontaneous Raman signal is limited to 1 dimensional multiplexing such as line excitation. Of course, spatially multiplexed signal generation requires light sources with proportionally scaled power.

Given the much slower acquisition speeds for spontaneous Raman scattering, it is difficult to find an obvious application in the histopathology work flow. For example, pixel sizes of 27 μm and a 90 Hz spectral acquisition rate achieved through spatial multiplexing would provide appropriate sample throughput, but with such course resolution, one would give up significant benefit of sub-cellular chemical detail otherwise available from Raman scattering. Spontaneous Raman imaging could be useful in a guided subsampling mode, where only regions of special interest were imaged. On the other hand, even with spatial multiplexing and 90 Hz spectral acquisition, only 0.003% of the total tissue area could be imaged at high resolution in the 2 minutes time frame required to obtain a 1 cm² whole slide brightfield image.

At this point it is appropriate to consider that the 2 minutes per slide imaging time is the standard for single-contrast image formation, whereas Raman spectral images could contain most or all of the potentially desired contrasts in a single, label-free image. Nonetheless, it seems clear that some trade-off between coverage and resolution would be necessary if spontaneous Raman scattering is used. The severity of such a trade-off could be reduced by use of coherent Raman techniques, which provide equivalent spectroscopic information with imaging throughput that is closer to that of WSI.

4 Coherent Raman imaging

4.1 Introduction and mechanism

Coherent Raman scattering was first predicted in 1962¹⁷¹ and first measured in 1964.¹⁷² In 1982, Duncan et al.¹⁷³ first demonstrated the coherent Raman scattering effect in a microscope, but the cross-beam (phase-matching) optical arrangement they used was challenging. In 1999 Zumbusch et al.¹⁹ demonstrated that the CARS effect could be achieved with a simple, collinear beam geometry. This discovery gave rise to the now burgeoning field of coherent Raman imaging (CRI). Currently, the most prominent bioimaging CRI methods are stimulated Raman scattering (SRS) and coherent anti-Stokes Raman scattering (CARS). It is not our purpose to review the CRI field in its entirety; excellent reviews have been given recently.^174–178 Instead, we focus on applications of CRI to histopathology, and associated technical aspects that facilitate or hinder such applications.

We can gain insight into coherent Raman scattering by considering that an oscillating electric field of light induces a polarization in a substance that can be described as:

P(t) = Nα(t)E sin(ωt) (1)

where P is an induced polarization, N is the number of molecules in the sample volume, α is the molecular polarizability, and E is the electric field amplitude of the light. As with spontaneous Raman scattering, the coherent Raman cross-section depends on changes of α with position along the normal mode coordinate, Q. We consider α in terms of its Taylor expansion about Q = 0.where n is the expansion order and m indexes normal modes.

In coherent scattering, vibrational modes are driven by external light fields. When the difference in energy between a pump (p) and Stokes (S) field is equal to the energy difference (Ω_v) between two adjacent quantum levels of the vibrational mode, i.e., when Ω_m = Δω = ω_p − ω_S, energy can be transferred to that vibrational mode. While the primary fields, E(t) = E_je^−iω_jt, (j = p,S), ω_p > ω_S ≫ Ω_m, are at much too high a frequency to directly induce significant molecular response, electrons are able to follow the fields adiabatically. Through its nonlinearity, the electron response will contain some amplitude at the difference frequency, Δω. This component of the response will in turn exert an oscillatory force F(t) on the molecule:¹⁷⁹

which will drive motion along the coordinate Q that can be described by:where Ω_m is the resonant frequency, μ is the reduced mass of the vibrational mode (not to be confused with the absorption dipole described previously), and γ is a damping constant. Q(t) is then given as:

Substituting eqn (5) and a first-order truncation of eqn (2) into eqn (1), we obtain an expression for the polarizability of a given mode in presence of a pump and Stokes field:

and the resonant susceptibility is given by the following sum over normal modes:where, by convention, N₀ is the number density of molecular oscillators in a pure substance. We see that χ⁽³⁾_R is complex and dissipative, so will have an overall spectral phase that varies with Ω. The imaginary component of χ⁽³⁾_R has essentially identical¹⁸⁰ line shape to the spontaneous Raman response. Molecular systems have an additional, nonresonant, susceptibility, χ⁽³⁾_NR, that arises directly through nonlinearity in the electronic response to the driving laser fields. The overall third order susceptibility is then given by:

χ⁽³⁾(Δω) = χ⁽³⁾_NR(Δω) + χ⁽³⁾_R(Δω) (8)

In third order coherent Raman processes, a pump and Stokes field interact with a medium to create a vibrational coherence, given as:

C(Δω) = χ⁽³⁾(Δω)[E_S ★ E_p](Δω) (9)

where E_p, E_S are the frequency-domain pump, and Stokes fields, respectively and ★ is the cross-correlation operator.

4.2 Coherent anti-stokes Raman scattering (CARS)

Coherent anti-Stokes Raman scattering is generated by interaction of a probe field with the vibrational coherence generated through the process described by eqn (9), as represented by the Jablonski diagram in Fig. 4a, and by the following expression:where E_pr is the frequency-domain probe field, ‘*’ is the convolution operator. For pump, Stokes, and probe sources with center frequencies of ω_p, ω_S, and ω_pr, respectively, the output anti-Stokes frequency ω_as = ω = ω_pr + ω_p − ω_S = ω_pr + Δω. In the limit that the probe pulse is spectrally narrow, we can treat it as a delta function, so the CARS signal contains the same spectral information that is contained in the spontaneous Raman signal. However, there are important differences between CARS and spontaneous Raman scattering. One of these is that the CARS signal is blue-shifted with respect to the excitation light, so does not compete with autofluorescence from the sample. Also, the CARS signal is forward-scattered, and contained in a narrow cone angle, making it a simple matter to collect a large fraction of the CARS light in a transmission geometry.

Fig. 4 (a) Abstraction of an experimental diagram for IR absorption or spontaneous Raman imaging: a single IR or visible light source is projected through an objective. For IR, absorption is measured from the relative light intensity transmitted through the sample. In spontaneous Raman scattering, signal is emitted isotropically, equally in the forward and backwards direction, but the Raman scattered light must be separated from the excitation light. This is typically done with spectral filters or spectrometer. Frequently, IR and spontaneous Raman imaging on tissue slides will be performed in reflection mode by placing a mirrored surface after the sample. The mirror serves to increase collected signal fraction in the case of spontaneous Raman scattering. (b) and (c): Energy diagrams for IR absorption and spontaneous Raman scattering respectively. (d) Abstraction of an experimental diagram for coherent Raman imaging: typically two pulsed and synchronized laser beams are co-linearly projected into the back aperture of the microscope objective. The CARS and SRS processes occur simultaneously at the focus. All of the SRS signal and most of the CARS signal are projected forward. A small CARS signal is projected in the backwards (epi) direction. SRS imaging is frequently performed in epi-detected mode by collecting signal light that is inelastically back-scattered into the excitation objective. The SRS signal is detected as a modulation of the pump or Stokes beam, while the CARS signal is spectrally distinct from the excitation light, and can be detected after spectral separation. (e) and (f): Energy diagrams for SRS and CARS respectively.

Another important difference is that CARS signal contains a nonresonant background (NRB), arising from χ⁽³⁾_NR. It is termed “nonresonant” because it arises from the electronic response, and is not sensitive to vibrational resonances. Since the electrons adiabatically follow the driving field of the laser, the NRB is always in phase with the driving field, and retains a fixed phase relationship with the resonant response. The overall CARS signal has a component that is a coherent mixing between these two terms:

Owing to the three terms in eqn (11), the analyte concentration dependence of the raw CARS signal depends on the ratio of resonant to nonresonant response. χ⁽³⁾_NR is of similar magnitude for most biological materials but χ⁽³⁾_R may vary widely owing to the density of molecular oscillators, Ñ (fraction of N₀ within the focal volume) and the Raman scattering cross-section of individual vibrational modes. Typically, χ⁽³⁾_NR ≤ 0.1Ñχ⁽³⁾_R for CH stretch resonances and I_CARS ∝ Ѳ. However, for many fingerprint peaks, χ⁽³⁾_NR ≪ Ñχ⁽³⁾_R and the resonant contribution to the I_CARS is linear in Ñ, but NRB dominates the signal. This is why narrowband CARS is performed almost exclusively with CH stretch resonances as contrast. In fact, considering only the term |Ñχ⁽³⁾_R|², Cui et al.¹⁸¹ demonstrated that the resonant CARS signal amplitude is greater than spontaneous Raman only when oscillators are very concentrated, such as in bulk phase, and samples rich in lipid or structural protein.

Another important effect of the coherent interaction between resonant and non-resonant signal components arises from the Gouy phase shift, wherein a phase shift accrues with field propagation beginning at the focus of a Gaussian beam compared to propagation of a plane wave. This becomes a problem at sharp interfaces of features smaller than the Rayleigh range when resonant signal is generated only by light on one side of the focus while the non-resonant signal is generated at positions on both sides of the focus. When these mix, there is a spatial position-dependent spectral phase error¹⁸² causing spectral distortion. However, the Gouy phase is not a problem when χ⁽³⁾_NR ≪ Ñχ⁽³⁾_R, or when there are no sharp spatial discontinuities in resonant signal contribution.¹⁸³

4.2.1 CARS with CH-stretch contrast. In the CH-stretch spectral range there are four main peaks, located at 2850 cm⁻¹, 2880 cm⁻¹, 2920 cm⁻¹, and 2960 cm⁻¹. These can have significant overlap, and correspond to symmetric and assymetric stretching of CH₂ (ν_CH2,_sym, ν_CH2,_asym), and CH₃ (ν_CH3,_sym, ν_CH3,_asym), respectively. There are also minor peaks at 3008 cm⁻¹ and 3060 cm⁻¹, due to ν_=CH and aromatic ν_CH and a broad resonance between 3000 cm⁻¹ and 3600 cm⁻¹ due to water ν_OH.

The CH stretch vibrations are sensitive to their environment by direct coupling and through Fermi resonances with C–H bending modes near 1500 cm⁻¹. Together, these influences can introduce significant shifts and broadening of the CH-stretch peaks. Spectroscopic^184–187 and spectral focusing¹⁸⁸ CARS approaches are best positioned to leverage the information contained in peak shape and position changes. Narrowband approaches¹⁹ return vibrational contrast over a narrow spectral band. In the CH stretch region, excitation is tuned to report primarily on a single resonance, but the signal generally includes lesser contributions from multiple resonances due to spectral overlap. Signal from CH₂ and CH₃ are frequently used to determine relative abundance of protein and lipid because the ratio of CH₂ and CH₃ moieties differs significantly in these chemical families. The information content in (symmetric, asymmetric) peak pairs is largely redundant, so only one peak from each pair is typically used for this purpose.

In narrowband CARS, a pair of picosecond pulses is used to excite coherence of a single selected vibrational band, and the higher frequency field is used to probe the coherence. Meyer et al.¹⁸⁹ found that, for strong resonant signals (χ⁽³⁾_R ≥ 10Nχ⁽³⁾_NR), simple subtraction of NRB is sufficient to render a quantitative signal with no significant distortions to the CARS peak positions and amplitudes. Wu et al.¹⁹⁰ have shown quantitative correspondence between ν_CH2,_sym signals and lipid content in liver tissue. Further, large resonant signals in the CH-stretch enables rapid image acquisition, with pixel rates as high as 6 MHz,¹⁹¹ but more typically range from 200 kHz (ref. 192) to 1 MHz (ref. 189) for tissue samples. Repeated images can be acquired at several discrete Raman shifts to map scenes with some chemical complexity. For discriminating species that are spectrally distinct, like water and protein or lipid, a single additional scan is generally sufficient for each component.¹⁹³ A larger number of spectral contrast points may be necessary to discriminate species with only subtle differences.¹⁸⁹ Given the high pixel acquisition rates, the necessity of acquiring several scans may not itself be onerous, but latency between scans for wavelength tuning can add significant time to image acquisition for most CARS systems. Rapid spectral scan approaches have been demonstrated for narrowband CARS with only 100 μs required to scan from one frequency to another in the CH stretch region, making inter-scan latency a non-issue for such systems.¹⁹⁴

Spectral focusing is another approach that facilitates spectroscopic scanning with narrowband CARS. In spectral focusing CARS,¹⁸⁸ a pair of pulses that are spectrally broad and centered at distinct wavelengths are temporally broadened with nominally identical chirp. Although the absolute frequencies of the two light fields impinging on the sample changes significantly with time, variation in the difference of frequencies is minimized so a single vibrational state can be excited. Using this approach, it is feasible to collect single-frequency image pixels at 250 kHz.¹⁹⁵ This approach to CARS can be performed with sub-picosecond pulses and spectral tuning over a limited range (typically 500 cm⁻¹) can be accomplished by simply changing the delay of one pulse relative to the other. When spectral focusing or another CARS approach is performed at a series of Raman shifts, the effects of the NRB (distortion and nonlinear concentration response) can be separated through spectral phase retrieval approaches.^196–198

Spectroscopic CARS differs from narrowband or spectral focusing CARS in that a broad swath of spectrum is obtained in each laser shot. Such an approach in the CH-stretch region has been demonstrated with picosecond/femtosecond CARS^{184,186,187,199} and with nonlinear interferometric vibrational imaging (NIVI).²⁰⁰ Histological studies focused on CH-stretch using spectroscopic CARS have been performed primarily with NIVI. From tissues, NIVI typically yields CH-stretch spectra with a bandwidth of ≈250 cm⁻¹ at a 20 Hz spectral acquisition rate (1 kHz effective rate for a single frequency assuming 5 cm⁻¹ resolution). These are undistorted Raman spectra, linear in N.²⁰⁰

Several studies have evaluated the utility of CARS in the CH stretch region for generating contrast equivalent to H&E stain. Evans et al.²⁰¹ compared CARS images to H&E stained images, finding that CARS could provide similar image information when ν_CH2,_sym and ν_CH3,_sym or ν_CH3,_asym were used. Similar “pseudo H&E” obtained through intravital CARS imaging²⁰² enabled inspection of tumor environments and margins without staining, and no obvious evidence of photodamage even after 300 frames.

Quantitative tissue morphology metrics can be obtained from CARS contrast, similar to what is done with H&E-contrast images. Meyer et al.²⁰³ compare CARS with spontaneous Raman in brain, demonstrating ability to extract nucleus-to-cytoplasm ratio, cell density, nucleus size and shape. They found, however, that since nuclei lack ν_CH2,_sym contrast, nuclear density and shape metrics may not be as reliable as H&E. On the other hand, the Wong group have used primarily ν_CH2,_sym CARS contrast and image segmentation^204–206 to determine nuclear size, cell volume, and cell–cell distance and other metrics to classify cancer subtypes in lung and breast tissue, producing encouraging results for differential diagnosis of cancer in these tissues. Uckermann et al.²⁰⁷ also compared metrics from CARS and H&E, finding that CARS provided additional relevant information via the signal amplitude.

In addition to conveying morphological information through imitating H&E stain, CARS ν_CH2 and ν_CH3 contrast has been used to study lipid metabolism in cancer,²⁰⁸ and to discriminate cell type based on lipid content. Excess lipids were found as a potential marker for circulating tumor cells,²⁰⁹ and Evans et al.²⁰¹ found contrast in ν_CH2,_sym to evince replacement of normal white matter with lipid deficient astrocytic glioma tissue. Chowdary et al.²¹⁰ found spectral differences arising from variation in lipid to protein ratio in a rat breast cancer model, and were able to use SVD to automatically segment normal and diseased tissue regions at a 99% confidence level. Similarly, Uckermann et al.²⁰⁷ also found a ν_CH2,_sym contrast gradient in the infiltrative zone around glioblastoma, with rather distinct contrast levels in the normal region. They also found that levels of ν_CH2,_sym contrast could be used to distinguish between glioblastoma, and metasticized breast tumor in the brain. In a similar finding, Meyer et al.¹¹⁹ showed that CARS ν_CH2,_sym contrast in the brain could highlight metastatic tumor regions of lung origin as the metastases retain a chemical profile similar to the originating tissue. Their finding was supported by Raman and IR imaging.

While narrowband CARS contrast from ν_CH2 and ν_CH3 alone is useful in many circumstances, it can also be combined synergistically with contrast from other methods, as exemplified in Fig. 5a and b. In other examples, Yue et al. used CARS signal from ν_CH3,_sym to visualize mammary acini, providing guidance for acquisition of spontaneous Raman spectra. From subsequent spectral analysis, they observe changes in spatial distribution of lipids concomitant with a loss of basoapical polarity in mammary acuinus, a transformation linked to early stages of carcinogens.¹⁷⁷ The Popp group have taken similar approaches.^{119,170,189,211} In comparing CARS to spontaneous Raman imaging of colon sections, Krafft et al.²¹² found that CARS images were consistent with expectations from spontaneous Raman scattering in the same spatial sample regions, and that with just a few CARS scans in the range 3000 to 1000 cm⁻¹, there was sufficient information to discriminate a couple of cell types and identify regions of normal and diseased tissue. The CARS images were acquired 10⁶ times faster than spontaneous Raman (10 μs vs. 30 s per pixel). Subsequently, they have used additional CARS peaks, in conjunction with other contrast mechanisms to characterize tissues. The same pulses that generate the CARS signal can also give rise to second harmonic generation (SHG) in collagen fibrils, and to two-photon excited fluorescence (TPEF) from intrinsic fluorophores such as NADPH, flavins, collagen, and elastin. Meyer et al.^119,189 used CARS contrast at multiple Raman shifts along with intrinsic contrast from TPEF and SHG to detect white matter, the granule layer, elastin, ordered collagen, and NAD(P)H in brain, and elastic fibers, triglycerides, collagen, myelin, cellular cytoplasm, and lipid droplets in perivascular tissue.

Fig. 5 a & b:²²¹ (a) Immunohistochemical stain for Sox-10 in mouse ear with haematoxylin counterstain, revealing melanocytes in red. (b) CARS image of adjacent slice to (a), showing pigmented regions, consistent with positive staining in (a). c & d:²¹⁸ (c) Pseudocolour BCARS image of tumor and normal brain tissue, with nuclei highlighted in blue, lipid content in red and red blood cells in green. (d) Raman spectra retrieved from (c), with each spectrum taken from one pixel, and acquired in 3.5 ms. e & f:²²² SRH image (e) is compared to a similar section of tumor imaged after formalin-fixation, paraffin-embedding and H&E staining (f). Comparison was made using a lookup table to match SRH “staining” with H&E. g & h:²²³ (g) Spectroscopic SRS images of pancreas and liver (left and right respectively). (h) Independent component spectra retrieved from (g) with 99% data reduction.

4.2.2 CARS with fingerprint contrast. As we have seen above, CH-stretch contrast is useful for morphology and some species quantification. However, contrast from fingerprint spectra can provide abundance information for many more species of diagnostic value. Some examples are shown in Table 2, including carotenoids, cholesterol esters, glycogen, collagen, keratin and elastin. Additionally, fingerprint spectra can positively identify cell phenotype.^162,168

Because the fingerprint peaks are typically much weaker than the nonresonant signal, it is not possible to obtain quantitative spectral information by simply subtracting the NRB. Because typically χ⁽³⁾_R ≪ χ⁽³⁾_NR for fingerprint resonances, the third (cross) term in eqn (11) dominates the resonant contribution to the CARS signal, and the overall spectral phase must be retrieved in order to extract a Raman spectrum. CARS phase retrieval is possible with an approach based on maximum entropy methods¹⁹⁶ or one based on time-domain Kramers–Kronig.¹⁹⁷ Both provide functionally equivalent results, but the computations for the Kramers–Kronig approach require considerably less time.²¹³ It has recently been appreciated that the NRB shape estimate required by both of these methods can have a profound impact on extracted Raman peak amplitudes. Camp et al.¹⁹⁸ showed that symmetry relations between the NRB and resonant signal can be used to unambiguously determine the NRB shape, and thus the relative peak amplitudes of the retrieved Raman spectrum. Additionally, the retrieved signal is linear with analyte concentration. In this way, the NRB acts as an internal calibration for spectroscopic CARS, and using this internal calibration, retrieved peak ratios are absolute.

Spectral acquisition of CARS signal is required for recovery of Raman fingerprint signals through phase retrieval, and also facilitates system noise removal via singular value decomposition.²¹⁴ Spectrally focused CARS¹⁹⁵ can be delay-scanned used to generate spectroscopic images that are amenable to phase retrieval²¹⁵ Additionally, several inherently spectroscopic CARS microscopy approaches have been developed, but most are too inefficient for biological imaging. Here we will exclusively consider approaches that have been used to generate images from cells and tissues. The first of these was based on a combination of spectrally broad but uncompressed Stokes pulses, and spectrally narrow probe pulses.^186,187 This approach generates broadband CARS spectra with up to 3000 cm⁻¹ bandwidth at ≈20 Hz spectral acquisition rate (12 kHz single spectral element, assuming 5 cm⁻¹ resolution), with strong CH stretch but quite weak fingerprint spectra.²¹⁶ The quality of fingerprint spectra was improved significantly in one case by implementation of a narrow-bandwidth sub-nanosecond probe,²¹⁷ and in another by using a temporally compressed Stokes source.²¹⁸ In the latter work, a nearly transform-limited broadband pulse acts as both pump and Stokes, exciting vibrational coherence much more efficiently than separate pulses. This approach yielded CARS spectra in both CH and fingerprint regions with uniformly high SNR across the spectrum, and 280 Hz spectral acquisition rate (corresponding to 170 kHz single spectral element acquisition, assuming 5 cm⁻¹ resolution).²¹⁸

Perhaps because it has only recently become possible to acquire high quality fingerprint spectra in an imaging modality, there is very little work on histological samples using fingerprint CARS. With relatively weak fingerprint signal, Pohling²¹⁹ was able to use principal component analysis to discriminate between gray and white matter, and identify layers of granule and Purkinje cells in mouse brain tissue. Their structural assignments agreed with those obtained from H&E staining. Also in murine brain tissue, but with significantly better SNR in the fingerprint due to impulsive coherence generation, Camp et al.^198,218 were able to identify many diagnostically important species and features directly from spectral peaks, without multivariate analysis techniques (Fig. 5c and d). These included red blood cells through the 1548 cm⁻¹ and 1565 cm⁻¹ hemoglobin peaks (C–C stretch), as well as collagen, elastin, and nucleotides through peaks listed in Table 2 for each of these species. They were able to identify tumor margins, as there were spectral differences between normal and invasive glioblastoma image regions, such as at 785 cm⁻¹ (nucleotide), 1004 cm⁻¹ (phenylalanine ring breathing), and 2956 cm⁻¹ (CH₃ stretch). In both the Pohling and Camp work, a tumor-associated reduction in ν_CH2 was observed, as in CARS studies that focused on CH-stretch.

In the mouse glioblastoma work, Camp et al.²¹⁸ obtained SHG and TPEF signals simultaneously with fingerprint CARS spectra. Although SHG is typically used to identify collagen,^119,220 the multi-peak spectral identification of this protein did not demonstrate identical results. They noted that collagen generates an SHG response only when the triple-helical structure is in-tact and only when the optical polarization and fibril axes are aligned.

4.3 Stimulated Raman scattering (SRS)

In stimulated Raman scattering (SRS), energy transfers from a higher frequency (pump) field to a lower frequency (Stokes) field when the frequency differences between the fields is equal to that of a vibrational resonance in the material being probed. This effect can be monitored as stimulated Raman loss (SRL) in the pump field, or stimulated Raman gain (SRG) in the Stokes field. An expression for SRL can be given as:

This expression differs from eqn (10) by addition of a pump field term. The pump field interacts coherently with the nonlinear polarization in the sample, generating a heterodyne term in the overall signal. Due to the relative phases of E_p and the field arising from the induced coherence, the heterodyne signal amplitude is proportional to only the imaginary part of χ⁽³⁾. Furthermore, only this heterodyne component is retrieved if the SRL is detected through modulation techniques, and one directly obtains a signal proportional to the spontaneous Raman scattering cross section (see eqn (6)). Because χ⁽³⁾_NR is real, the nonresonant component that is problematic for narrowband CARS is not detected in SRS. In spite of this, the theoretically achievable signal-to-noise ratio for SRS is the same as that of CARS,²²⁴ and a related value, the sensitivity limit, seems to be similar for both approaches.²¹⁸ The heterodyne component of the SRL signal can be expressed as:

For field strengths permissible in tissue microscopy, the relative change in pump or Stokes intensity is on the order of 10⁻⁶ to 10⁻⁸, even for the strong peaks in the CH stretch region. These small signals cannot be spectrally separated from the intense excitation light as in CARS, so must be discriminated by other means, such as lock-in or balanced detection methods. Under favorable conditions, pixel acquisition rates up to 7 MHz have been demonstrated with contrast from Raman peaks in the CH stretch region.²²⁵ However, 30–40 kHz pixel rates are more common for single spectral bands in tissue samples.^222,226 Furthermore, even the strongest fingerprint peaks, such as found at 1440 & 1660 cm⁻¹ (CH₂ deformation & C=C stretch) are challenging to use for image contrast with SRS. Ozeki et al.²²⁷ were unable to directly use the 1660 cm⁻¹ peak for contrast, even when pixel dwell times were 10× longer than necessary for good SNR in the CH-stretch region.

While the NRB does not interfere with the SRS signal, other background signals can. These include two photon absorption (TPA), cross phase modulation and thermal lensing. Signals from these sources can have amplitude comparable to or greater than even the largest SRS signals. For example, in skin, the TPA signal exceeds that of CH stretch SRS by almost 100-fold.²²¹ Not surprisingly, these background signals can easily mask weak contrast from the fingerprint region, and careful measures may be taken to suppress their effects and recover some of the stronger fingerprint signals.^228,229

Because the SRS signal must be extracted directly from the laser field, lasers with very low residual intensity noise (RIN) are preferred. These are typically solid state lasers. However, fiber lasers have a number of qualities that make them desirable in a clinical setting, including being robust to physical movement, vibration, and temperature shifts, and requiring very little maintenance. A low-noise fiber source has been developed specifically for SRS²³⁰ although even for this laser, it is necessary to use both balanced detection and lock-in approaches to recover the SRS signal.

4.3.1 SRS with CH-stretch contrast. As with CARS, SRS imaging in the CH-stretch region has been performed with several approaches, including dual narrowband pulses,^231,232 spectral focusing,²³³ rapid spectral scanning,^227,234 and simultaneous multi-band imaging.²³⁵ The same chemical information can be obtained with CARS and SRS in the CH-stretch region. Accordingly, the types of histology studies with SRS are quite similar to those of CARS.

Using only contrast from the ν_CH3,_sym band and morphological clues, Mittal et al.²³⁶ were able to identify cell membranes, nuclei, collagen and keratin in tissues containing squamous cell carcinoma. In that study, the authors found that SRS images from this single band provided similar information to that found in H&E images.

Lu et al.²³⁷ compared images from H&E staining to those generated with SRS contrast from ν_CH2,_sym and ν_CH3,_sym. They found that cell density counts from SRS and H&E images were indistinguishable. However, using the lipid/protein ratio from the SRS signal, and confirming this with luxol fast blue staining for the myelin sheath, they found that the two SRS bands provided a better distinction between white and gray matter than could be achieved with H&E staining. They were also able to identify red blood cells and vessels through morphological features of images with contrast at 2800 cm⁻¹, in the nominally quiescent region. The contrast there may have arisen from TPA of heme. They also found that necrotic regions were characterized by lack of nuclei and lower protein signal, and banded regions of high protein signal appeared to be collagen.

Ji et al.²³⁸ used SRS contrast for ν_CH2,_symm and ν_CH3,_symm in fresh surgical brain tissues from 22 patients to create a classifier based on cellularity, axnoal density, and protein to lipid ratios, which provided tumor infiltration detection with 97.5% sensitivity and 98.5% specificity. They found that SRS images were superior to H&E images for detecting infiltrating glioma, and that SRS microscopy could detect invasive tumor cells in peri-tumor brain tissue that appears grossly normal.

Orringer et al.²²² used ν_CH2,_symm and ν_CH3,_symm bands from SRS to recapitulate image contrast from H&E, but on unprocessed tissue, with an eye towards improving intrasurgical tissue sampling. Example images from that study are shown in Fig. 5e and f. Pathologist diagnosis based on proximally located SRS and H&E tissue samples from 30 patients was compared and found to agree 92% of the time. They also used computer algorithm-derived image attributes (based on SRS contrast) to successfully distinguish between diagnostic classes of tumors with 90% accuracy.

Cui et al.²³⁹ used SRS spectra in the region (2800 to 3000) cm⁻¹ and multivariate curve resolution to find a correlation between cholesterol ester content in intracellular lipid droplets and Gleason scores of prostate tissue. Using acalibrated surrogate mixture for cellular lipid content, they found that the ratio between a band assigned to cholesterol at 2870 cm⁻¹ and the ν_CH2,_sym band was proportional to the fraction of total lipids due to cholesterol ester.

While full spectral information in the CH-stretch region is sometimes useful, a small number of peaks often provide sufficient discrimination. Otsuka et al.²²³ examined tissues from a tumor-graft mouse model using spectrally scanned SRS spectrally scanned over the range (2800 to 3100) cm⁻¹. They found that through independent component analysis, they could reduce their spectral dataset by >95%, and still obtain images in which nucleus, zymogen granules, lipid droplets, and red blood cells could be visualized, and irregular glandular cell alignments detected (see Fig. 5g and h). They found that images from the reduced spectral data was qualitatively equivalent to H&E stained images. Similarly, Egawa et al.²⁴⁰ used SRS spectroscopy in the range (2800 to 3100) cm⁻¹ to map spatially resolved keratinocyte differentiation, and distinguish these from Langerhans cells in the spinous layer of the epidermis. The key differences were primarily morphological, as both cell shape and nuclear prominence differ between the cell types, so the important contrast came at ν_CH2,_symm and ν_CH3,_symm bands.

In other studies, vibrational signal from SRS has been used in concert with other nonlinear contrast mechanisms. Freudiger et al.²²⁶ combined ν_CH2,_sym and ν_CH3,_sym bands with hemoglobin TPA to derive contrast of roughly equal quality to H&E-stained sections, recapitulating major histologic features in tissue from several brain regions. Galli et al.²⁴¹ used ν_CH2 in conjunction with TPEF of an extrinsic label, and SHG from structured collagen to observe significant differences in nuclear area of normal and cancerous regions in human tumors and murine tumor models. Yue et al.²⁴² used SRS ν_CH2,_sym contrast for morphological imaging in liver and prostate tissue, and for guidance in identifying lipid droplets (LDs). They then used spontaneous Raman spectroscopy at LD positions to determine that the cholesterol ester content of the LDs was correlated with the ratio of 702 and 1442 cm⁻¹ peak intensity ratios, and with the aggressiveness of the cancer.

4.3.2 SRS with fingerprint contrast. As with CARS, little histology CRI has been performed with contrast in the fingerprint region. Ozeki²²⁷ showed that, for typical modulation schemes, it is difficult to acquire image contrast using fingerprint Raman peaks. In CARS, the problem of weak fingerprint signal was overcome with a highly efficient signal generation mechanism.²¹⁸ While no analogous enhanced-efficiency signal generation approach has been found for SRS, modulation schemes have been demonstrated which discriminate against parasitic background signals sufficiently that fingerprint peaks can be observed. Berto et al.²²⁸ proposed a scheme that exploits the even modulation symmetry of various background signals, providing sufficient discrimination to acquire fingerprint spectra from mouse skin at a 20 Hz spectral acquisition rate (720 Hz spectral pixel). Zhang et al.²⁴³ were able to image cell nuclei using SRS with contrast from DNA at 785 cm⁻¹, at pixel acquisition rates on the order of 10 kHz. Fu et al.²²⁹ demonstrated an SRS modulation approach that appears to 3× to 4× more effective for rejecting background signals that vary slowly with excitation wavelength than amplitude modulation. They used delay-modulated spectral focusing achieve ≈10 mmol L⁻¹ sensitivity from 720 cm⁻¹ peak of acetylcholine in neuronal tissue with a 250 kHz pixel acquisition rate.

Colon tissues from 9 patients were imaged by spectrally scanned SRS in conjunction with SHG, THG and TPEF.²⁴⁴ They found that an increase in protein (ν_CH3,_sym and Amide I), and water (3250 cm⁻¹), and a decrease in lipid (ν_CH2,_sym, 1445 cm⁻¹, 1745 cm⁻¹) all correlated with presence of adenocarcinoma. They reported similarity in TPEF of NADH and 1665 cm⁻¹ contrast images, suggesting that increased protein could be due to increased metabolic activity. They also suggested that increased water in cancerous regions could be due to increased aquaporin protein production.

5 A future role for coherent Raman imaging in histopathology

We expect that CRI methods for histopathology will be adopted to the extent that they cost-effectively meet specific needs, and fit well into the established histology work flow, possibly streamlining it. As early applications are explored, and as physicians become increasingly comfortable with the technology, it is possible that more complex and potentially disruptive adaptations of this technology may be adopted. Of course, initial adoption will be influenced by the costs of CRI instrumentation relativeto existing technology. In this section, we briefly discuss classes of problems that CRI can address, both in the near and longer terms.

5.1 Immediacy of information

It is not uncommon to have 24 hours latency between tissue excision and histology analysis. While this is often acceptable, there are situations in which immediate characterization is optimal or even required, such as while sampling a suspected tumor, and during an intra-operative margin evaluation. In these cases, label-free and in situ histology modalities are desirable.

Tissue involved in and surrounding a tumor is typically sampled and analyzed by histology or for biomarkers at various time points to guide therapy by establishing prognosis and following disease progression. Accordingly, inappropriate tumor sampling is thought to be a leading problem in both prognosis and evaluating treatment response.²⁴⁵

Tissue sampling by needle biopsy or exploratory surgery is often guided by palpitation, preparative imaging, or visual inspection, and is notoriously inefficient.²⁴⁵ Neoplasms may be difficult to properly sample at early stages, and the ability to categorize tissue as normal or neoplastic during the biopsy process would go a long way towards ensuring high-quality samples for subsequent analysis. Spontaneous Raman was demonstrated decades ago as an potential guidance tool for needle biopsy,²⁴⁶ and progress has been made toward clinical adaptation, but challenges remain with regards to low signal levels, and typically, Raman needle probes must be coupled with other detection modalities.¹¹³ Coherent Raman methods could easily provide sufficient signal levels for this application. Narrowband CRI approaches have been implemented in fiber,^247–249 so could be adapted to needle biopsy, but are probably not suited for this application since disease detection would be based strictly on spectra, and thus benefit greatly from sensitivity to multiple peaks in the fingerprint spectral region. Current broadband spectroscopic CRI approaches based on impulsive excitation or spectral focusing will be difficult to implement in a fiber delivery, but other approaches may be possible in the future.

While narrowband CRI may not be appropriate for purely spectroscopic disease detection, in needle biopsy, it is effective in a microscopy modality for in vivo detection and analysis of skin carcinomas, which are present in the optically accessible epidermis layer of the skin.^{211,221,236,250} New approaches to signal detection may push the useful range of in vivo CRI to slightly deeper tissues,²⁵¹ and to detection of fingerprint signals.²²⁹

Another widely recognized clinical problem is the need for intraoperative analysis of tissue at a resection margin. Strong correlations exist between presence of cancer cells in surgical margins and recurrence rates of cancer,^252,253 and incomplete margin resection is held as a major cause for recurrence of cancers that call for tissue conservation, such as in brain and breast.

A hand-held probe is required for in situ margin analysis, and appropriate equipment has already been described for narrowband CRI^247–249 although, to our knowledge, clinical studies using these probes have not yet been published. Use of a hand-held Raman probe for identification of tumor tissue in surgical margins was clinically demonstrated for brain surgery by Jermyn et al.²⁵⁴ and proof of principle for breast was demonstrated by Kong et al.²⁵⁵ As with Raman needle probes, tumor detection in these reports was based on spectroscopy alone, thus suffering somewhat from low signal and precluding imaging. This difficulty was avoided in the breast application by using autofluorescence levels to identify regions of interest, and those regions were imaged sparsely with Raman. In the brain application, the surgeon visually selected regions of interest.

Although CRI technology has not yet progressed to in situ margin analysis, it has been demonstrated as an effective tool for characterizing freshly excised, unprocessed tissue.²²² Typically, margin tissue must be frozen and stained for imaging. This process can add 20 minutes to a surgery each time a margin is to be investigated. The label-free nature of CRI, and the ability to image in a backscattering mode makes it possible to obtain histology-quality images from surface regions of thick, unprocessed tissue, significantly reducing the delay between tissue removal and analysis.²²²

Another area in which speed of information availability would be useful is in initial evaluation of histology slides. The majority of tissue in most slides is normal, so a pathologist may spend a significant portion of their time searching for potential diseased regions. One particularly strong benefit of chemical imaging is that it can provide clear indications of tissue type and regions of normal and potentially abnormal tissue. This regional tissue typing has already been demonstrated with IR on slides from which high-resolution H&E images have been obtained, and the two images overlaid.¹⁶¹ CRI methods can provide both the regional tissue typing^201,207,210 and the high-resolution H&E mimic without the need for staining or imaging in multiple contrast modes.^201,202,222

5.2 Better quantification of morphological parameters

A survey of pathologists by the National Collation on Health Care and Best Doctors³⁶ gave lack of sub-specialty expertise the most often suspected reason for cancer misdiagnoses from histology. In that study, physicians most frequently indicated that improved pathology resources (i.e., innovations in histopathology) is the thing most needed to improve diagnostic accuracy. One particular approach to improving pathology resources, namely CAD, aims to bridge the sub-specialty expertise gap.

An important example of how CAD can bridge this gap is through providing basic quantification of image features. As mentioned previously, non-specialist pathologists were found to be much less precise in cell counts and classification than sub-specialists,⁸⁸ no doubt contributing significantly to overall diagnostic uncertainty. Many of the same factors contributing to non-specialist variability, including the fact that cells are often overlaid in H&E images, with no clear delineation between nuclei,⁸⁸ are also challenges to the families of morphological image analysis on which CAD is based. Further, factors such as variability in staining protocols, and in image field illumination, and issues with nonspecific staining contribute to overall image variability, adversely impacting reliability in thresholding and feature identification for CAD. These, and related issues are expected to be significant barriers to widespread adoption of CAD. In recognition of this, the US Food and Drug Administration has issued guidelines for performance of WSI instrumentation.⁵⁷

CRI methods can ameliorate or resolve many of the reproducibility and reliability issues in image formation. For example, obtaining reliable cell counts will be easier because CRI has intrinsically high axial resolution, making it feasible to resolve stacked nuclei. Additionally, where imaged nuclei are spatially close, there will be a nuclear and cellular membrane signal between them, strengthening the ability of CRI-based analysis to arrive at accurate cell counts. Yang et al.²⁰⁶ have already demonstrated 90% accuracy and reproducibility in automated cell counting using coherent Raman imaging with only ν_CH2,_sym, and thus poor nuclear contrast.²⁰³ This is an important near-term opportunity for coherent Raman methods.

In addition to intrinsic z-sectioning, CRI can provide a large number of contrasts through various vibrational bands. Since increased image dimensionality generally leads to higher precision segmentation and feature identification,^89–91 one could expect improved performance in these aspects with these multi-contrast images. The Wong group have derived morphological parameters familiar to pathologists, such as cell shape and density, from CARS images and based on these were able to identify with 80 to 95% accuracy breast tumor types.²⁰⁶

Although CRI requires little or no sample preparation, it seems likely that most early implementations will include fixation and paraffin embedding of samples, simply for consistency with protocols for staining. In fact, both of these steps have potential to introduce variability in CRI signals. Paraffin has a strong CH stretch signal, and fixing agents can liberate diagnostically important molecules, particularly lipids.²⁵⁶ For example, compared to fresh frozen tissue, formalin-fixed brain showed modification in ν_CH2,_assym and ν_CH3,_assym amplitudes from CARS, but not in ν_CH2,_sym. In the fingerprint, only choline groups of phospholipids at 717 cm⁻¹ dropped slightly after formalin fixing. However, methanol or acetone fixation severely reduced all of these peaks and others.²⁵⁷

5.3 Increased chemical information

Morphological analysis of an image from a single stain such as H&E is sufficient for diagnosis in most cases. However, difficult cases often require multiple additional stains to provide needed auxiliary information, but many of these desirable stains are rare or unavailable,² and it seems that this lack of access is a clinically significant issue.³⁶ CRI with good fingerprint contrast has potential to make an important contribution by providing a considerable range of contrasts.

Fingerprint CRI provides chemical specificity identical to that of spontaneous Raman spectroscopy,²¹⁸ and therefore seems likely to be a potential substitute for most or all stains, and perhaps many IHC agents. Raman spectra can be used to recapitulate many image contrast agents used commonly in biology.²⁵⁸ Also, as shown by the partial listing in Table 2, fingerprint Raman has already been used to detect many diagnostic markers. With fingerprint CRI, these will be available as label-free image contrasts. An important added benefit of spectroscopic CARS is that it provides an internal reference based on NRB,¹⁹⁸ so that spectroscopic CARS images can serve as quantitative, rather than qualitative maps of markers. This level of quantification cannot generally be achieved with stains, fluorescence, or even spontaneous Raman. The quantitative imaging and absence of non-specific staining will greatly simplify image analysis and increase confidence in image metrics and associated diagnostic decisions. A few specific examples of potential important contributions are given below.

The number density of mitotic cells plays an important role in diagnosis and grading of many cancer types,⁸ and is usually determined by IHC or cell morphology analysis. The best morphologically-derived automated detection of mitoses achieve F-scores (the geometric mean of precision and accuracy) between 75% and 80%.⁸⁷ It seems likely that fingerprint CRI imaging could yield significantly better results. By mapping the intranuclear distribution of DNA, RNA, and chromatin, Raman spectroscopic imaging can not only positively identify whether a cell is mitotic, but which stage of mitosis it is in.^163,166,167 The chemical distribution information is essentially orthogonal to morphological analysis, so can only serve to increase precision and accuracy in determining mitosis counts.

Accurate characterization of pleomorphism (atypical cell shape) is also important to proper diagnosis of many cancers.⁸ Pleomorphic cells of one type may be mistakenly identified as normal cells of a different type, and this was part of a 42% cell count and classification disagreement among non-specialists found by Fuchs et al.⁸⁸ It is well established that Raman spectroscopy can be used to positively identify cell type.¹⁶² Applying spectral cell identification as an orthogonal metric will facilitate much better recognition of pleomorphism.

Invasion of immune cells into a tumor site, and their state of activation is an key factor in diagnosis of some cancers.⁸ As mentioned above, morphology-based cell identification can be difficult, and determination of immune cell activation is currently done by assessing the presence of cell surface markers, typically by IHC. For this important problem, fingerprint spectroscopic imaging will not only allow immune cell identification, but also whether they are activated.^107,168

Identifying the tissue from which of CUP (cancer unknown primary) metastases originate is useful for identifying appropriate therapeutic strategies, but presently requires involved analysis.^259–261 Fingerprint spectroscopic CRI may be able to positively identify metastatic tumor origins simply from chemical content of the tumor cells. Even narrowband CRI studies were able to discriminate primary (metastasized) cancer cells from cells belonging to the target tissue simply on basis of contrast at ν_CH2,_symm and ν_CH3,_symm.^119,207

In recent years it has become clear that genomic subclasses of cancers can have distinct clinical outcomes,²⁶² and thus require varied interventions. While spectroscopic Raman imaging has not been evaluated as a tool for identifying these subclasses, it seems reasonable that it may have power to do so in some cases. CRI will not directly detection genome changes, but it is likely to detect downstream consequences of those changes. For example, cell phenotype¹⁶² and function^107,162 can be discriminated using Raman spectra. These spectral changes come not from DNA sequence or methylation modifications directly, but from the intra- and extra-cellular changes in bimolecular profiles they induce. Furthermore, microenvironment properties determined through histology have recently been found to correlate with the genomic cancer subtypes, and these microenvironment changes should be detectable with CRI in most cases.²⁶³

5.3.1 Speed & information trade-off. Single frequency CRI approaches can provide imaging nearly as fast as WSI instrumentation. CARS approaches typically provide tissue imaging at pixel rates of 200 kHz to 1 MHz,^189,192 as little as a factor of three slower than commercial WSI. SRS instruments seem to run a bit more slowly, typically at ≈50 kHz.^222,226 However, a recent SRS demonstration achieves fingerprint contrast at 250 kHz, between 10 and 60 times slower than WSI. Although these speeds may still be too low for widespread use, they present only a minimal imaging rate-related barrier to applying narrowband CRI to histopathology problems in a research mode. Accordingly, reports of narrowband histology studies are beginning to appear. On the other hand, many of CRI's potential contributions to histopathology will require chemical information beyond that attainable from a single Raman peak. Based on current technology, there will be a significant speed penalty for increased chemical information.

Broadband CRI tissue imaging using CH stretch has been demonstrated at 20 Hz for CARS¹⁸⁵ and at 4 kHz for SRS^223,240 over 300 cm⁻¹. Fingerprint CARS spectro-microscopy has been demonstrated at 300 Hz over 3000 cm⁻¹ in tissue.²¹⁸ Overall, these range in speed from 10⁵ to 10³ times slower than WSI. This difference may be difficult to surmount. Assuming the 0.5 cm² min⁻¹ imaging speed is indeed an appropriate mark, it seems that present spectral CRI methods are not good candidates, although some newer spectroscopic CRI approaches may hold promise. A spectral focusing approach has been demonstrated at 77 kHz for a 200 cm⁻¹ bandwidth,²⁵¹ although not yet in tissue, and a family of approaches featuring ≈25 kHz spectroscopic CARS acquisition has been demonstrated,^264,265 but with laser pulses incompatible with tissue imaging. Broadband SRS^266,267 has been demonstrated with potentially fast imaging, but seem to suffer from insufficient modulation sensitivity.

Assuming that signal generation and collection efficiency of current CRI instruments is nearly optimal, and there is no way to intrinsically speed up image acquisition, there are at least two non-exclusive options for applying spectroscopic CRI to histopathology. One is to use a faster imaging modality (such as narrowband CRI) to generate a guide image from which regions of interest can be determined, and use spectroscopic CRI over ROIs covering a small fraction of the specimen. The other approach would be to use spatial multiplexing to effectively increase the rate of spectroscopic signal acquisition.

Provided sufficient excitation laser power, spatial multiplexing approaches can yield speed increases proportional to the number of detection elements. For spectroscopically detected signals, such as in spontaneous Raman scattering, 1-D (line) spatial multiplexing can realistically lead to 20×–100× imaging speed increase.¹⁶⁴ This approach could also be used for spectroscopic CARS. 2-D multiplexing has been demonstrated with narrowband CARS^268–272 though not in ways suited for clinical applications. Spectroscopic approaches using single element detectors, such as time-domain²⁶⁴ or spectrally scanned¹⁹⁵ CARS could also be 2-D multiplexed, so considerable speed increases are conceivable, provided laser sources of sufficient pulse energy can be found to spread light over a larger area and still obtain signal. SRS will be more challenging to implement in spatially multiplexed modes since it requires very high sensitivity to small amplitude modulations. 2-D array detectors, such as CMOS cameras can be used in differential or lock-in mode, but these have yet to demonstrate sensitivity that would be needed for tissue imaging.^266,267 Slipchenko et al.²⁷³ introduced an array of 32 LC circuits for simultaneously detecting SRS signal in as many channels, which has shown sufficient sensitivity for tissue imaging.

5.3.2 Other considerations. Issues such as instrument costs and data storage requirements will also contribute to the probability that coherent Raman imaging would be adopted in clinical practice. While it is difficult to account for market factors, one might guess at a future product price by considering an existing instrument of similar complexity and function. Two-photon microscopes come to mind for such a comparison, since they require pulsed light sources similar to CRI instrumentation. Such microscopes have been engineered to be reasonably user-friendly, and are commonly deployed in biological research laboratories. A quality two-photon microscope can be obtained at a cost the order of $0.5 M, and this is in the ballpark of what one might expect for a high quality commercial WSI system.

Finally, we note that routine use of spectroscopic CRI in histopathology could present significant data storage issues. A single-contrast slide image from a WSI instrument may occupy 1–2 Gb of storage. At a rate of 500 images a day, storing digital histopathology data is already a challenge. If all such images were acquired in full spectral mode, the data storage load would increase by 10³ times – an untenable proposition. Such a scenario is neither likely or necessary; there are at least two factors that would mitigate the need for such excessive storage. One is that spectroscopic imaging would be particularly beneficial for only small areas of most tissues. The other is that, while spectral images could be stored temporarily, for processing, it quite unlikely that saving all spectral dimensions will be necessary in most cases.¹⁶⁹ Data reduction methods such as those discussed in Sections 2.7.3 and 3.1.1 could be applied to determine the most important spectroscopic information, allowing to discard the rest and reduce the size of images for long-term storage.

6 Conclusions

We expect that coherent Raman imaging is most likely to find acceptance in the clinic as a drop-in replacement for current technologies to the extent that it is able to help resolve recognized problems related to detection, diagnosis, and treatment of disease. Such problems are likely to include fast and reliable intra-surgical margin evaluation, either through direct analysis of the surgical site or through rapid analysis of freshly excised tissue. Similarly, CRI's ability to provide label-free molecularly specific signal quickly may be useful for in situ detection of small tumors or neoplasms in exploratory surgery or needle biopsy.

Intra-surgical microscopic analysis of excised tissue has already been demonstrated. In situ use will likely require a hand-held probe with fiber-delivered excitation light and signal collection. Whether in situ tissue analysis will be best done by narrowband coherent Raman imaging or by coherent Raman spectroscopic analysis is unclear at this time.

Another significant opportunity for CRI is facilitating improvements in histology-based diagnoses. The art of histopathology emerged near the end of the 19^th century in nearly its present form. Significant new knowledge of disease mechanisms has improved interpretation of histological features, but, apart from the introduction of IHC in the 1940s, histopathology tools in use today are largely the same as they were in the early 20^th century. Efforts to modernize histopathology practice are making headway, however, fully transforming it from a subjective art to a quantitative and objective science remains a challenge. Most of these challenges involve variability and contrast reliability in image formation. It appears that many of these issues, could be largely resolved with CRI. Furthermore, spectroscopic CRI may provide a significantly enhanced contrast space that will open new options for quickly stratifying cancer and other diseases, leading to a more personalized diagnosis.

While opportunities for CRI to improve clinical practice are clear, important barriers remain for widespread adoption in histology practice. An increase in imaging speed for narrowband CARS and a slightly larger increase for SRS would be helpful. For spectroscopic CRI, substantial imaging speed increases will be necessary if they are to be used to a significant extent in an imaging mode. Fortunately, the CRI field is still vibrant with new innovations for increased imaging speed and chemical selectivity and sensitivity. Much work remains to demonstrate the utility and reliability of these methods to the medical community. Finally, potential regulatory issues must be considered and resolved to the extent that they are found.

Conflicts of interest

There are no conflicts to declare.

Acknowledgements

We thank C. Paul Morris, M.D., Ph.D. for helpful discussions and comments on the manuscript.

References

1. D. Evanko, A. Heinrichs and C. K. Rosenthal, Nat. Cell Biol., 2009, 11, S6–S22.
2. M. Titford, J. Histotechnol., 2009, 32, 9–19.
3. G. Hunninghake, M. B. Zimmerman, D. Schwartz, T. King, J. Lynch, R. Hegele, J. Waldron, T. Colby, N. Muller, D. Lynch, J. Galvin, B. Gross, J. Hogg, G. Toews, R. Helmers, J. A. Cooper, R. Baugman, C. Strange and M. Millard, Am. J. Respir. Crit. Care Med., 2001, 164, 193–196.
4. B. W. G. van Rhijn, G. J. L. H. van Leenders, B. C. M. Ooms, W. J. Kirkels, A. R. Zlotta, E. R. Boevé and A. C. Jöbsis, Eur. Urol., 2010, 57, 1052–1057.
5. W. C. Allsbrook Jr., K. A. Mangold, M. H. Johnson, R. B. Lane, C. G. Lane and J. I. Epstein, Hum. Pathol., 2001, 32, 81–88.
6. M. Costantini, S. Sciallero, A. Giannini, B. Gatteschi, P. Rinaldi, G. Lanzanova, L. Bonelli, T. Casetti, E. Bertinelli, O. Giuliani, G. Castiglione, P. Mantellini, C. Naldoni and P. Bruzzi, J. Clin. Epidemiol., 2003, 56, 209–214.
7. T. A. Longacre, M. Ennis, L. A. Quenneville, A. L. Bane, I. J. Bleiweiss, B. A. Carter, E. Catelano, M. R. Hendrickson, H. Hibshoosh, L. J. Layfield, L. Memeo, H. Wu and F. P. O'Malley, Mod. Pathol., 2005, 19, 195–207.
8. Stanford Medicine: Surgical Pathology Criteria, http://www.surgpathcriteria.stanford.edu.
9. A. H. Coons, H. J. Creech and R. N. Jones, Exp. Biol. Med., 1941, 47, 200–202.
10. C. D. Fletcher, J. J. Berman, C. Corless, F. Gorstein, J. Lasota, B. J. Longley, M. Miettinen, T. J. O'Leary, H. Remotti, B. P. Rubin, B. Shmookler, L. H. Sobin and S. W. Weiss, Hum. Pathol., 2002, 33, 459–465.
11. M. Werner, A. Chott, A. Fabiano and H. Battifora, Am. J. Surg. Pathol., 2000, 24, 1016.
12. M. Sabah, M. Leader and E. Kay, Appl. Immunohistochem. Mol. Morphol., 2003, 11, 56–61.
13. T. Kirkegaard, J. Edwards, S. Tovey, L. M. McGlynn, S. N. Krishna, R. Mukherjee, L. Tam, A. F. Munro, B. Dunne and J. M. S. Bartlett, Histopathology, 2006, 48, 787–794.
14. R. S. Weinstein, A. R. Graham, L. C. Richter, G. P. Barker, E. A. Krupinski, A. M. Lopez, K. A. Erps, A. K. Bhattacharyya, Y. Yagi and J. R. Gilbertson, Hum. Pathol., 2009, 40, 1057–1069.
15. D. L. Woernley, Cancer Res., 1952, 12, 516–523.
16. C. Kendall, N. Stone, N. Shepherd, K. Geboes, B. Warren, R. Bennett and H. Barr, J. Pathol., 2003, 200, 602–609.
17. R. Bhargava and I. W. Levin, Appl. Spectrosc., 2004, 58, 995–1000.
18. A. S. Haka, K. E. Shafer-Peltier, M. Fitzmaurice, J. Crowe, R. R. Dasari and M. S. Feld, Proc. Natl. Acad. Sci. U. S. A., 2005, 102, 12371.
19. A. Zumbusch, G. R. Holtom and X. S. Xie, Phys. Rev. Lett., 1999, 82, 4142–4145.
20. M. J. Nasse, M. J. Walsh, E. C. Mattson, R. Reininger, A. Kajdacsy-Balla, V. Macias, R. Bhargava and C. J. Hirschmugl, Nat. Methods, 2011, 8, 413–416.
21. K. Yeh, S. Kenkel, J.-N. N. Liu and R. Bhargava, Anal. Chem., 2015, 87, 485–493.
22. Protein Atlas, http://www.proteinatlas.org.
23. W. H. Fridman, F. Pagès, C. Sautès-Fridman and J. Galon, Nat. Rev. Cancer, 2012, 12, 298–306.
24. W. Schubert, B. Bonnekoh, A. J. Pommer, L. Philipsen, R. Böckelmann, Y. Malykh, H. Gollnick, M. Friedenberger, M. Bode and A. W. M. Dress, Nat. Biotechnol., 2006, 24, 1270–1278.
25. P. A. Humphrey, J. Clin. Pathol., 2007, 60, 35–42.
26. J. D. Kronz, W. H. Westra and J. I. Epstein, Cancer, 1999, 86, 2426–2435.
27. S. S. Raab, R. E. Nakhleh and S. G. Ruby, Arch. Pathol. Lab. Med., 2005, 129, 459–466.
28. J. Ho, A. V. Parwani, D. M. Jukic, Y. Yagi, L. Anthony and J. R. Gilbertson, Hum. Pathol., 2006, 37, 322–331.
29. A. A. Renshaw, K. M. Lezon and D. C. Wilbur, Cancer Cytopathol., 2001, 93, 106–110.
30. A. Lurkin, F. Ducimetière, D. R. Vince, A.-V. Decouvelaere, D. Cellier, F. N. Gilly, D. Salameire, P. Biron, G. de Laroche, J. Y. Blay and I. Ray-Coquard, BMC Cancer, 2010, 10, 150.
31. J. M. Bruner, L. Inouye, G. N. Fuller and L. A. Langford, Cancer, 1997, 79, 796–803.
32. P. L. Nguyen, D. Schultz, A. A. Renshaw, R. T. Vollmer, W. R. Welch, K. Cote, A. V. DAmico and A. V. D'Amico, Urol. Oncol.: Semin. Orig. Invest., 2004, 22, 295–299.
33. V. L. Staradub, Ann. Surg. Oncol., 2002, 9, 982–987.
34. T. R. Coblentz, S. E. Mills and D. Theodorescu, Cancer, 2001, 91, 1284–1290.
35. A. E. Selman, T. H. Niemann, J. M. Fowler and L. J. Copeland, Obstet. Gynecol., 1999, 94, 302–306.
36. MisdiagnosisSurvey_FINALiv.pdf, http://www.bestdoctors.com/about-best-doctors/news-and-events/nchc-misdiagnosis-survey.
37. S. C. Hollensead, W. B. Lockwood and R. J. Elin, J. Surg. Oncol., 2004, 88, 161–181.
38. M. M. Mello, A. Chandra, A. A. Gawande and D. M. Studdert, Health Aff., 2010, 29, 1569–1577.
39. ACS_2016_FactsFigures.pdf, http://www.cancer.org.
40. Cost of Cancer Care, 2012.
41. Z. Huang, A. McWilliams, H. Lui, D. I. McLean, S. Lam and H. Zeng, Int. J. Cancer, 2003, 107, 1047–1052.
42. N. Stone, C. Kendall, J. Smith, P. Crow and H. Barr, Faraday Discuss., 2004, 126, 141–157.
43. T. C. Bakker Schut, K. Maquelin, T. van der Kwast, C. H. Bangma, D.-J. J. Kok and G. J. Puppels, Anal. Chem., 2006, 78, 7761–7769.
44. M. Gniadecka, P. A. Philipsen, S. Sigurdsson, S. Wessel, O. F. Nielsen, D. H. Christensen, J. Hercogova, K. Rossen, H. K. Thomsen, R. Gniadecki, L. K. Hansen and H. C. Wulf, J. Invest. Dermatol., 2004, 122, 443–449.
45. M. Uhlén, E. Björling, C. Agaton, C. A.-K. Szigyarto, B. Amini, E. Andersen, A.-C. C. Andersson, P. Angelidou, A. Asplund, C. Asplund, L. Berglund, K. Bergström, H. Brumer, D. Cerjan, M. Ekström, A. Elobeid, C. Eriksson, L. Fagerberg, R. Falk, J. Fall, M. Forsberg, M. G. Björklund, K. Gumbel, A. Halimi, I. Hallin, C. Hamsten, M. Hansson, M. Hedhammar, G. Hercules, C. Kampf, K. Larsson, M. Lindskog, W. Lodewyckx, J. Lund, J. Lundeberg, K. Magnusson, E. Malm, P. Nilsson, J. Odling, P. Oksvold, I. Olsson, E. Oster, J. Ottosson, L. Paavilainen, A. Persson, R. Rimini, J. Rockberg, M. Runeson, A. Sivertsson, A. Sköllermo, J. Steen, M. Stenvall, F. Sterky, S. Strömberg, M. A. Sundberg, H. Tegel, S. Tourle, E. Wahlund, A. Waldén, J. Wan, H. Wernérus, J. Westberg, K. Wester, U. Wrethagen, L. L. Xu, S. Hober and F. Pontén, Mol. Cell. Proteomics, 2005, 4, 1920–1932.
46. J. B. Welsh, L. M. Sapinoso, S. G. Kern, D. A. Brown, T. Liu, A. R. Bauskin, R. L. Ward, N. J. Hawkins, D. I. Quinn, P. J. Russell, R. L. Sutherland, S. N. Breit, C. A. Moskaluk, H. F. Frierson and G. M. Hampton, Proc. Natl. Acad. Sci. U. S. A., 2003, 100, 3410–3415.
47. K. Honda, T. Yamada, R. Endo, Y. Ino, M. Gotoh, H. Tsuda, Y. Yamada, H. Chiba and S. Hirohashi, J. Cell Biol., 1998, 140, 1383–1393.
48. P. Mauri, A. Scarpa, A. C. Nascimbeni, L. Benazzi, E. Parmagnani, A. Mafficini, M. D. Peruta, C. Bassi, K. Miyazaki and C. Sorio, FASEB J., 2005, 19, 1125–1127.
49. P. Alfonso, A. Núñez, J. Madoz-Gurpide, L. Lombardia, L. Sánchez and J. I. Casal, Proteomics, 2005, 5, 2602–2611.
50. R. M. Levenson and J. R. Mansfield, Cytometry, Part A, 2006, 69, 748–758.
51. P. Zrazhevskiy, L. D. True and X. Gao, Nat. Protoc., 2013, 8, 1852–1869.
52. M. J. Gerdes, C. J. Sevinsky, A. Sood, S. Adak, M. O. Bello, A. Bordwell, A. Can, A. Corwin, S. Dinn, R. J. Filkins, D. Hollman, V. Kamath, S. Kaanumalle, K. Kenny, M. Larsen, M. Lazare, Q. Li, C. Lowes, C. C. McCulloch, E. McDonough, M. C. Montalto, Z. Pang, J. Rittscher, A. Santamaria-Pang, B. D. Sarachan, M. L. Seel, A. Seppo, K. Shaikh, Y. Sui, J. Zhang and F. Ginty, Proc. Natl. Acad. Sci. U. S. A., 2013, 110, 11982–11987.
53. D. P. Riordan, S. Varma, R. B. West and P. O. Brown, PLoS One, 2015, 10, e0128975.
54. E. Schubert, W. Gross, R. H. Siderits, L. Deckenbaugh, F. He and M. J. Becich, Semin. Diagn. Pathol., 1994, 11, 263–273.
55. L. Pantanowitz, N. Farahani and A. Parwani, Pathol. Lab. Med. Int., 2015, 23.
56. L. Pantanowitz, P. N. Valenstein, A. J. Evans, K. J. Kaplan, J. D. Pfeifer, D. C. Wilbur, L. C. Collins and T. J. Colgan, J. Pathol. Inform., 2011, 2, 36.
57. Guidance for Industry and Food and Drug Administration Staff, http://fda.gov/downloads/MedicalDevices/DeviceRegulationandGuidance/GuidanceDocuments/UCM435355.pdf.
58. A. J. Méndez, P. G. Tahoces, M. J. Lado, M. Souto and J. J. Vidal, Med. Phys., 1998, 25, 957–964.
59. M. Orcutt, MIT Technology Review, 2015.
60. L. He, L. R. Long, S. Antani and G. Thoma, Seq. Genome Anal., 2010, 510, 271–287.
61. H. Mousavi, V. Monga, G. Rao and A. Rao, J. Pathol. Inform., 2015, 6, 15–15.
62. J. Barker, A. Hoogi, A. Depeursinge and D. L. Rubin, Med. Image Anal., 2016, 30, 60–71.
63. E. Bengtsson, H. Danielsen, D. Treanor, M. N. Gurcan, C. MacAulay and B. Molnár, Cytometry, Part A, 2017, 91, 551–554.
64. M. N. Gurcan, L. E. Boucheron, A. Can, A. Madabhushi, N. M. Rajpoot and B. Yener, IEEE Rev. Biomed. Eng., 2009, 2, 147–171.
65. H. Irshad, A. Veillard, L. Roux and D. Racoceanu, IEEE Rev. Biomed. Eng., 2014, 7, 97–114.
66. G. D. Marty, BioTechniques, 2007, 42, 716.
67. A. C. Ruifrok and D. A. Johnston, Anal. Quant. Cytol. Histol., 2001, 23, 291–299.
68. M. Macenko, M. Niethammer, J. S. Marron, D. Borland, J. T. Woosley, X. Guan, C. Schmitt and N. E. Thomas, Biomedical Imaging: From Nano to Macro, 2009. ISBI'09. IEEE International Symposium on, 2009, pp. 1107–1110.
69. A. M. Khan, N. Rajpoot, D. Treanor and D. Magee, IEEE Trans. Biomed. Eng., 2014, 61, 1729–1738.
70. N. Otsu, Automatica, 1975, 11, 23–27.
71. S. G. Chang, B. Yu and M. Vetterli, IEEE Trans. Image Proc., 2000, 9, 1522–1531.
72. J. B. Roerdink and A. Meijster, Fundam. Inform., 2000, 41, 187–228.
73. J. C. Bezdek, R. Ehrlich and W. Full, Comput. Geosci., 1984, 10, 191–203.
74. D. L. Pham, C. Y. Xu and J. L. Prince, Ann. Rev. Biomed. Eng., 2000, 2, 315–337.
75. T. K. Moon, IEEE Signal Processing Magazine, 1996, 13, 47–60.
76. M. Kass, A. Witkin and D. Terzopoulos, Int. J. Comput. Vis., 1988, 1, 321–331.
77. J. S. Suri, K. Liu, S. Singh, S. N. Laxminarayan, X. Zeng and L. Reden, IEEE Trans. Inform. Technol. Biomed., 2002, 6, 8–28.
78. S. Naik, S. Doyle, S. Agner, A. Madabhushi, M. Feldman and J. Tomaszewski, 2008 5th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2008, pp. 284–287.
79. M. Alilou, V. Kovalev and V. Taimouri, Comput. Med. Imaging. Graph., 2013, 37, 488–499.
80. I. Jolliffe, Principal component analysis, Springer, New York, 2nd edn, 2002.
81. I. T. Joliffe and B. J. T. Morgan, Stat. Methods Med. Res., 1992, 1, 69–95.
82. A. Hyvärinen, J. Karhunen and E. Oja, Independent component analysis, John Wiley & Sons, New York, 2001, vol. 46.
83. A. M. Martínez and A. C. Kak, IEEE Trans. Pattern Anal. Mach. Intell., 2001, 23, 228–233.
84. S. Doyle, M. Hwang, K. Shah, A. Madabhushi, M. Feldman and J. Tomaszeweski, Biomedical imaging: from nano to macro, 2007. ISBI 2007. 4th IEEE international symposium on, 2007, pp. 1284–1287.
85. L. Breiman, Mach. Learn., 2001, 45, 5–32.
86. M. Valkonen, K. Kartasalo, K. Liimatainen, M. Nykter, L. Latonen and P. Ruusuvuori, Cytometry, Part A, 2017, 91, 555–565.
87. D. C. Cirean, A. Giusti, L. M. Gambardella and J. Schmidhuber, Medical Image Computing and Computer-Assisted Intervention MICCAI 2013: 16th International Conference, Nagoya, Japan, September 22–26, 2013, Proceedings, Part II, Springer Berlin Heidelberg, Berlin, Heidelberg, 2013, pp. 411–418.
88. T. J. Fuchs and J. M. Buhmann, Comput. Med. Imaging. Graph., 2011, 35, 515–530.
89. D. C. Fernandez, R. Bhargava, S. M. Hewitt and I. W. Levin, Nat. Biotechnol., 2005, 23, 469–474.
90. S. S. Singh, B. Qaqish, J. L. Johnson, O. H. Ford 3rd, J. F. Foley, S. J. Maygarden and J. L. Mohler, Anal. Quant. Cytol. Histol., 2004, 26, 194–200.
91. D. L. Weaver, D. N. Krag, E. A. Manna, T. Ashikaga, S. P. Harlow and K. D. Bauer, Mod. Pathol., 2003, 16, 1159–1163.
92. M. E. Brezinski, G. J. Tearney, B. E. Bouma, J. A. Izatt, M. R. Hee, E. A. Swanson, J. F. Southern and J. G. Fujimoto, Circulation, 1996, 93, 1206.
93. Z. Wang, K. Tangella, A. Balla and G. Popescu, J. Biomed. Opt., 2011, 16, 116017.
94. Y. L. Kim, Y. Liu, R. K. Wali, H. K. Roy, M. J. Goldberg, A. K. Kromin, K. Chen and V. Backman, IEEE J. Sel. Top. Quantum Electron., 2003, 9, 243–256.
95. H. Majeed, S. Sridharan, M. Mir, L. Ma, E. Min, J. Opt. Soc. Am. BW. Jung and G. Popescu, J. Biophotonics, 2017, 10, 177–205.
96. N. Lue, J. Bewersdorf, M. D. Lessard, K. Badizadegan, R. R. Dasari, M. S. Feld and G. Popescu, Opt. Lett., 2007, 32, 3522–3524.
97. P. Wang, R. Bista, R. Bhargava, R. E. Brand and Y. Liu, Opt. Lett., 2010, 35, 2840–2842.
98. V. Backman, M. B. Wallace, L. T. Perelman, J. T. Arendt, R. Gurjar, M. G. Müller, Q. Zhang, G. Zonios, E. Kline and T. McGillican, Nature, 2000, 406, 35.
99. Z. Wang, H. Ding and G. Popescu, Opt. Lett., 2011, 36, 1215–1217.
100. R. Richards-Kortum and E. Sevick-Muraca, Annu. Rev. Phys. Chem., 1996, 47, 555–606.
101. R. Alfano, D. Tata, J. Cordero, P. Tomashefsky, F. Longo and M. Alfano, IEEE J. Quantum Electron., 1984, 20, 1507–1511.
102. K. T. Schomacker, J. K. Frisoli, C. C. Compton, T. J. Flotte, J. M. Richter, N. S. Nishioka and T. F. Deutsch, Lasers Surg. Med., 1992, 12, 63–78.
103. S. G. Demos, R. Gandour-Edwards, R. Ramsamooj and R. d. White, J. Biomed. Opt., 2004, 9, 587–592.
104. M. Monici, Biotechnol. Annu. Rev., 2005, 11, 227–256.
105. B. Lin, S. Urayama, R. M. Saroufeem, D. L. Matthews and S. G. Demos, Opt. Express, 2010, 18, 21074–21082.
106. K. Kong, C. J. Rowlands, S. Varma, W. Perkins, I. H. Leach, A. A. Koloydenko, H. C. Williams and I. Notingher, Proc. Natl. Acad. Sci. U. S. A., 2013, 110, 15189–15194.
107. M. Diem, A. Mazur, K. Lenau, J. Schubert, B. Bird, M. Miljković, C. Krafft and J. Popp, J. Biophotonics, 2013, 6, 855–886.
108. M. J. Baker, J. Trevisan, P. Bassan, R. Bhargava, H. J. Butler, K. M. Dorling, P. R. Fielden, S. W. Fogarty, N. J. Fullwood, K. A. Heys, C. Hughes, P. Lasch, P. L. Martin-Hirsch, B. Obinaju, G. D. Sockalingum, J. Sulé-Suso, R. J. Strong, M. J. Walsh, B. R. Wood, P. Gardner and F. L. Martin, Nat. Protoc., 2014, 9, 1771–1791.
109. Q. Matthews, A. Brolo, J. Lum, X. Duan and A. Jirasek, Phys. Med. Biol., 2011, 56, 19–38.
110. C. Krafft, S. B. Sobottka, G. Schackert and R. Salzer, J. Raman Spectrosc., 2006, 37, 367–375.
111. L. A. Austin, S. Osseiran and C. L. Evans, Analyst, 2016, 141, 476–503.
112. K. Eberhardt, C. Stiebing, C. Matthäus, M. Schmitt and J. Popp, Expert Rev. Mol. Diagn., 2015, 15, 773–787.
113. O. Stevens, I. E. Iping Petterson, J. C. C. Day and N. Stone, Chem. Soc. Rev., 2016, 45, 1919–1934.
114. E. Montgomery, M. P. Bronner, J. R. Goldblum, J. K. Greenson, M. M. Haber, J. Hart, L. W. Lamps, G. Y. Lauwers, A. J. Lazenby, D. N. Lewin, M. E. Robert, A. Y. Toledano, Y. Shyr and K. Washington, Hum. Pathol., 2001, 32, 368–378.
115. M. Jermyn, K. Mok, J. Desroches, J. Mercier, K. Saint-Arnaud, L. Bernstein, M.-C. C. Guiot, K. Petrecca and F. Leblond, Biomedical Optics, 2014, Miami, Florida, 2014, p. BS5A.4.
116. T. Hollon, S. Lewis, C. W. Freudiger, X. Sunney Xie and D. A. Orringer, Neurosurg. Focus, 2016, 40, E9.
117. R. Manoharan, Y. Wang and M. S. Feld, Spectrochim. Acta, Part A, 1996, 52, 215–249.
118. C. Petibois and G. Deleris, Trends Biotechnol., 2006, 24, 455–462.
119. T. Meyer, N. Bergner, C. Bielecki, C. Krafft, D. Akimov, B. F. M. Romeike, R. Reichart, R. Kalff, B. Dietzek and J. Popp, J. Biomed. Opt., 2011, 16, 21110–21113.
120. I. Pence and A. Mahadevan-Jansen, Chem. Soc. Rev., 2016, 45, 1958–1979.
121. D. Shipp, F. Sinjab and I. Notingher, Adv. Opt. Photonics, 2017, 9, 315–428.
122. K. Fukuda, Histochemie, 1966, 6, 127–130.
123. F. F. M. De Mul, H. Buiteveld, J. Lankester, J. Mud and J. Greve, Hum. Pathol., 1984, 15, 1062–1068.
124. N. Stone, P. Stavroulaki, C. Kendall, M. Birchall and H. Barr, Laryngoscope, 2000, 110, 1756–1763.
125. A. Beljebbar, O. Bouché, M. D. Diébold, P. J. Guillou, J. P. Palot, D. Eudes and M. Manfait, Crit. Rev. Oncol./Hematol., 2009, 72, 255–264.
126. M. Gniadecka, H. C. Wulf, O. F. Nielsen, D. H. Christensen and J. Hercogova, Photochem. Photobiol., 1997, 66, 418–423.
127. S. Koljenović, L.-P. Choo-Smith, T. C. Bakker Schut, J. M. Kros, H. J. van den Berge and G. J. Puppels, Lab. Invest., 2002, 82, 1265–1277.
128. S. Rehman, Z. Movasaghi, A. T. Tucker, S. P. Joel, J. A. Darr, A. V. Ruban and I. U. Rehman, J. Raman Spectrosc., 2007, 38, 1345–1351.
129. P.-H. H. Chen, R. Shimada, S. Yabumoto, H. Okajima, M. Ando, C.-T. T. Chang, L.-T. T. Lee, Y.-K. K. Wong, A. Chiou and H.-o. O. Hamaguchi, Sci. Rep., 2016, 6, 20097.
130. S. Kaminaka, H. Yamazaki, T. Ito, E. Kohda and H.-o. O. Hamaguchi, J. Raman Spectrosc., 2001, 32, 139–141.
131. A. Nijssen, K. Maquelin, L. F. Santos, P. J. Caspers, T. C. Bakker Schut, J. C. den Hollander, M. H. A. Neumann and G. J. Puppels, J. Biomed. Opt., 2007, 12, 034004.
132. A. Mizuno, H. Kitajima, K. Kawauchi, S. Muraishi and Y. Ozaki, J. Raman Spectrosc., 1994, 25, 25–29.
133. J. T. Kwak, A. Kajdacsy-Balla, V. Macias, M. Walsh, S. Sinha and R. Bhargava, Sci. Rep., 2015, 5, 8758.
134. R. Manoharan, K. Shafer, L. Perelman, J. Wu, K. Chen, G. Deinum, M. Fitzmaurice, J. Myles, J. Crowe, R. R. Dasarl and M. S. Feld, Photochem. Photobiol., 1998, 67, 15–22.
135. C. H. Liu, B. B. Das, W. L. S. Glassman, G. C. Tang, K. M. Yoo, H. R. Zhu, D. L. Akins, S. S. Lubicz, J. Cleary, R. Prudente, E. Celmer, A. Caron and R. R. Alfano, J. Photochem. Photobiol., B, 1992, 16, 187–209.
136. U. Utzinger, D. L. Heintzelman, A. Mahadevan-Jansen, A. Malpica, M. Follen and R. Richards-Kortum, Appl. Spectrosc., 2001, 55, 955–959.
137. S. You, H. Tu, Y. Zhao, Y. Liu, E. J. Chaney, M. Marjanovic and S. A. Boppart, Sci. Rep., 2016, 6, 32922.
138. R. E. Kast, G. K. Serhatkulu, A. Cao, A. K. Pandya, H. Dai, J. S. Thakur, V. M. Naik, R. Naik, M. D. Klein, G. W. Auner and R. Rabah, Biopolymers, 2008, 89, 235–241.
139. M. Miljković, T. Chernenko, M. J. Romeo, B. Bird, C. Matthäus and M. Diem, Analyst, 2010, 135, 2002–2013.
140. B. Vajna, G. Patyi, Z. Nagy, A. Bódis, A. Farkas and G. Marosi, J. Raman Spectrosc., 2011, 42, 1977–1986.
141. H. J. Butler, L. Ashton, B. Bird, G. Cinque, K. Curtis, K. Esmonde-white, N. J. Fullwood, B. Gardner, P. L. Martin, M. J. Walsh, M. R. Mcainsh, N. Stone, F. L. Martin, H. J. Butler and P. L. Martin-hirsch, Nat. Protoc., 2016, 11, 1–47.
142. F. Masia, A. Glen, P. Stephens, P. Borri and W. Langbein, Anal. Chem., 2013, 85, 10820–10828.
143. J. J. Andrew and T. M. Hancewicz, Appl. Spectrosc., 1998, 52, 797–807.
144. I. I. Patel, J. Trevisan, G. Evans, V. Llabjani, P. L. Martin-Hirsch, H. F. Stringfellow and F. L. Martin, Analyst, 2011, 136, 4950–4959.
145. D. Zhang, P. Wang, M. N. Slipchenko, D. Ben-Amotz, A. M. Weiner and J. X. Cheng, Anal. Chem., 2013, 85, 98–106.
146. A. Alfonso-García, J. Paugh, M. Farid, S. Garg, J. Jester and E. Potma, J. Raman Spectrosc., 2017, 803–812.
147. A. J. Berger, T. W. Koo, I. Itzkan, G. Horowitz and M. S. Feld, Appl. Opt., 1999, 38, 2916–2926.
148. M. S. Bergholt, W. Zheng, K. Y. Ho, M. Teh, K. G. Yeoh, J. B. Yan So, A. Shabbir and Z. Huang, Gastroenterology, 2014, 146, 27–32.
149. E. Widjaja, W. Zheng and Z. Huang, Int. J. Oncol., 2008, 32, 653–662.
150. P. Rösch, M. Harz, M. Schmitt, O. Ronneberger, H. Burkhardt, M. Lankers, S. Hofer, H. Thiele, P. Ro, K.-d. D. Peschke and H.-w. W. Motzkus, Appl. Environ. Microbiol., 2005, 71, 1626–1637.
151. S. K. Teh, W. Zheng, D. P. Lau and Z. Huang, Analyst, 2009, 134, 1232.
152. A. Kallenbach-Thieltges, F. Großerüschkamp, A. Mosig, M. Diem, A. Tannapfel and K. Gerwert, J. Biophotonics, 2013, 6, 88–100.
153. O. J. Old, G. R. Lloyd, J. Nallala, M. Isabelle, L. M. Almond, N. A. Shepherd, C. A. Kendall, A. C. Shore, H. Barr and N. Stone, Analyst, 2016, 1227–1234.
154. A. J. Berger, T.-w. Koo, I. Itzkan and M. S. Feld, Anal. Chem., 1998, 70, 623–627.
155. G. R. Lloyd, L. M. Almond, N. Stone, N. Shepherd, S. Sanders, J. Hutchings, H. Barr and C. Kendall, Analyst, 2014, 139, 381–388.
156. Y. LeCun, B. Yoshua and H. Geoffrey, Nature, 2015, 521, 436–444.
157. M. Gniadecka, H. C. Wulf, N. N. Mortensen, O. F. Nielsen and D. H. Christensen, J. Raman Spectrosc., 1997, 28, 125–129.
158. R. Goodacre, E. M. Timmins, R. Burton, N. Kaderbhai, A. M. Woodward, D. B. Kell and P. J. Rooney, Microbiology, 1998, 144, 1157–1170.
159. P. Manescu, Y. Jong Lee, C. Camp, M. Cicerone, M. Brady and P. Bajcsy, Med. Image Anal., 2017, 37, 37–45.
160. M. J. Pilling, A. Henderson, J. H. Shanks, M. D. Brown, N. W. Clarke and P. Gardner, Analyst, 2016, 1258–1268.
161. J. T. Kwak, S. M. Hewitt, S. Sinha and R. Bhargava, BMC Cancer, 2011, 11, 62.
162. J. W. Chan, D. K. Lieu, T. Huser and R. A. Li, Anal. Chem., 2009, 81, 1324–1331.
163. C. Matthäus, S. Boydston-White, M. Miljković, M. Romeo and M. Diem, Appl. Spectrosc., 2006, 60, 1–8.
164. M. Okada, N. I. Smith, A. F. Palonpon, H. Endo, S. Kawata, M. Sodeoka and K. Fujita, Proc. Natl. Acad. Sci. U. S. A., 2012, 109, 28–32.
165. H. G. Schulze, S. O. Konorov, J. M. Piret, M. W. Blades and R. F. Turner, Analyst, 2013, 138, 3416–3423.
166. S. Boydston-White, T. Chernenko, A. Regina, M. S. Miljković, C. Matthäus and M. Diem, Vib. Spectrosc., 2005, 38, 169–177.
167. S. Boydston-White, T. Gopen, S. Houser, J. Bargonetti and M. Diem, Biospectroscopy, 1999, 5, 219–227.
168. J. W. Chan, D. S. Taylor, S. M. Lane, T. Zwerdling, J. Tuscano and T. Huser, Anal. Chem., 2008, 80, 2180–2187.
169. S. Tiwari, J. Raman, V. Reddy, A. Ghetler, R. P. Tella, Y. Han, C. R. Moon, C. D. Hoke and R. Bhargava, Anal. Chem., 2016, 88, 10183–10190.
170. C. Krafft, M. Schmitt, I. W. Schie, D. Cialla-May, C. Matthäus, T. Bocklitz and J. Popp, Angew. Chem., Int. Ed., 2016, 4392–4430.
171. J. A. Armstrong, N. Bloembergen, J. Ducuing and P. S. Pershan, Phys. Rev., 1962, 127, 1918–1939.
172. P. D. Maker and R. W. Terhune, Phys. Rev., 1965, 137, A801–A818.
173. M. D. Duncan, J. Reintjes and T. J. Manuccia, Opt. Lett., 1982, 7, 350–352.
174. A. Volkmer, J. Phys. D: Appl. Phys., 2005, 38, R59.
175. C. L. Evans and X. S. Xie, Annu. Rev. Anal. Chem., 2008, 1, 883–909.
176. J. P. Pezacki, J. A. Blake, D. C. Danielson, D. C. Kennedy, R. K. Lyn and R. Singaravelu, Nat. Chem. Biol., 2011, 7, 137–145.
177. S. Yue, J. Cárdenas-Mora, L. Chaboub, S. Lelièvre and J.-X. X. Cheng, Biophys. J., 2012, 102, 1215–1223.
178. C. H. Camp Jr. and M. T. Cicerone, Nat. Photonics, 2015, 9, 295–305.
179. E. O. Potma and S. Mukamel, in Theory of Coherent Raman Scattering, CRC Press, Boca Raton, FL, 2013.
180. W. M. Tolles, J. W. Nibler, J. R. McDonald and A. B. Harvey, Appl. Spectrosc., 1977, 31, 253–271.
181. M. Cui, B. R. Bachler and J. P. Ogilvie, Opt. Lett., 2009, 34, 773–775.
182. K. I. Popov, A. F. Pegoraro, A. Stolow and L. Ramunno, Opt. Express, 2011, 19, 5902–5911.
183. A. M. Barlow, K. Popov, M. Andreana, D. J. Moffatt, A. Ridsdale, A. D. Slepkov, J. L. Harden, L. Ramunno and A. Stolow, Opt. Express, 2013, 21, 15298–15307.
184. J. X. Chen, A. Volkmer, L. D. Book and X. S. Xie, J. Phys. Chem. B, 2002, 106, 8493–8498.
185. D. L. Marks and S. A. Boppart, Phys. Rev. Lett., 2004, 92, 123905.
186. T. W. Kee and M. T. Cicerone, Opt. Lett., 2004, 29, 2701–2703.
187. H. Kano and H. Hamaguchi, Appl. Phys. Lett., 2004, 85, 4298–4300.
188. T. Hellerer, A. M. Enejder and A. Zumbusch, Appl. Phys. Lett., 2004, 85, 25–27.
189. T. Meyer, M. Chemnitz, M. Baumgartl, T. Gottschall, T. Pascher, C. Matthäus, B. F. M. Romeike, B. R. Brehm, J. Limpert, A. Tünnermann, M. Schmitt, B. Dietzek and J. Popp, Anal. Chem., 2013, 85, 6703–6715.
190. Y.-M. M. Wu, H.-C. C. Chen, W.-T. T. Chang, J.-W. W. Jhan, H.-L. L. Lin and I. Liau, Anal. Chem., 2009, 81, 1496–1504.
191. C. L. Evans, E. O. Potma, M. Puoris’ haag, D. Côté, C. P. Lin and X. S. Xie, Proc. Natl. Acad. Sci. U. S. A., 2005, 102, 16807–16812.
192. I. Veilleux, J. A. Spencer, D. P. Biss, D. Cote and C. P. Lin, IEEE J. Sel. Top. Quantum Electron., 2008, 14, 10–18.
193. E. Y. Lin, J. G. Jones, P. Li, L. Zhu, K. D. Whitney, W. J. Muller and J. W. Pollard, Am. J. Pathol., 2003, 163, 2113–2126.
194. S. Bégin, B. Burgoyne, V. Mercier, A. Villeneuve, R. Vallée and D. Côté, Biomed. Opt. Express, 2011, 2, 1296–1306.
195. A. F. Pegoraro, A. Ridsdale, D. J. Moffatt, J. P. Pezacki, B. K. Thomas, L. B. Fu, L. Dong, M. E. Fermann and A. Stolow, Opt. Express, 2009, 17, 20700–20706.
196. E. M. Vartiainen, J. Opt. Soc. Am. B, 1992, 9, 1209–1214.
197. Y. X. Liu, Y. J. Lee and M. T. Cicerone, Opt. Lett., 2009, 34, 1363–1365.
198. C. H. Camp, Y. J. Lee and M. T. Cicerone, J. Raman Spectrosc., 2016, 47, 408–415.
199. M. Muller and J. M. Schins, J. Phys. Chem. B, 2002, 106, 3715–3723.
200. W. A. Benalcazar, P. D. Chowdary, Z. Jiang, D. L. Marks, E. J. Chaney, M. Gruebele and S. A. Boppart, IEEE J. Sel. Top. Quant., 2010, 16, 824–832.
201. C. L. Evans, X. Xu, S. Kesari, X. S. Xie, S. T. Wong and G. S. Young, Opt. Express, 2007, 15, 12076–12087.
202. M. Lee, A. Downes, Y.-Y. Y. Chau, B. Serrels, N. Hastie, A. Elfick, V. Brunton, M. Frame and A. Serrels, IntraVital, 2015, e1055430.
203. T. Meyer, N. Bergner, A. Medyukhina, B. Dietzek, C. Krafft, B. F. M. Romeike, R. Reichart, R. Kalff and J. Popp, J. Biophotonics, 2012, 5, 729–733.
204. L. Gao, F. Li, M. J. Thrall, Y. Yang, J. Xing, A. A. Hammoudi, H. Zhao, Y. Massoud, P. T. Cagle and Y. Fan, J. Biomed. Opt., 2011, 16, 096004–096004.
205. L. Gao, Z. Wang, F. Li, A. A. Hammoudi, M. J. Thrall, P. T. Cagle and S. T. Wong, Arch. Pathol. Lab. Med., 2012, 136, 1502–1510.
206. Y. Yang, F. Li, L. Gao, Z. Wang, M. J. Thrall, S. S. Shen, K. K. Wong and S. T. Wong, Biomed. Opt. Express, 2011, 2, 2160–2174.
207. O. Uckermann, R. Galli, S. Tamosaityte, E. Leipnitz, K. D. Geiger, G. Schackert, E. Koch, G. Steiner and M. Kirsch, PLoS One, 2014, 9, e107115.
208. T. T. Le, T. B. Huff and J. X. Cheng, BMC Cancer, 2009, 9, 42.
209. R. Mitra, O. Chao, Y. Urasaki, O. B. Goodman and T. T. Le, BMC Cancer, 2012, 12, 540.
210. P. D. Chowdary, Z. Jiang, E. J. Chaney, W. A. Benalcazar, D. L. Marks, M. Gruebele and S. A. Boppart, Cancer Res., 2010, 70, 9562–9569.
211. S. Heuke, N. Vogler, T. Meyer, D. Akimov, F. Kluschke, H.-J. J. Röwert-Huber, J. Lademann, B. Dietzek and J. Popp, Br. J. Dermatol., 2013, 169, 794–803.
212. C. Krafft, A. A. Ramoji, C. Bielecki, N. Vogler, T. Meyer, D. Akimov, P. Rösch, M. Schmitt, B. Dietzek, I. Petersen, A. Stallmach and J. Popp, J. Biophoton., 2009, 2, 303–312.
213. M. T. Cicerone, K. A. Aamer, Y. J. Lee and E. Vartiainen, J. Raman Spectrosc., 2012, 43, 637–643.
214. A. Rajwade, A. Rangarajan and A. Banerjee, IEEE Trans. Pattern Anal. Mach. Intell., 2013, 35, 849–862.
215. A. F. Pegoraro, A. D. Slepkov, A. Ridsdale, D. J. Moffatt and A. Stolow, J. Biophotonics, 2012, 49–58.
216. S. H. Parekh, Y. J. Lee, K. A. Aamer and M. T. Cicerone, Biophys. J., 2010, 99, 2695–2704.
217. M. Okuno, H. Kano, P. Leproux, V. Couderc, J. P. R. Day, M. Bonn and H.-o. O. Hamaguchi, Angew. Chem., Int. Ed., 2010, 49, 6773–6777.
218. C. H. Camp Jr., Y. J. Lee, J. M. Heddleston, C. M. Hartshorn, W. R. Hight, J. N. Rich, J. D. Lathia and M. T. Cicerone, Nat. Photonics, 2014, 8, 627–634.
219. C. Pohling, T. Buckup, A. Pagenstecher and M. Motzkus, Biomed. Opt. Express, 2011, 2, 2110–2116.
220. Y. J. Lee, Y. Liu and M. T. Cicerone, Opt. Lett., 2007, 32, 3370–3372.
221. H. Wang, S. Osseiran, V. Igras, A. J. Nichols, E. M. Roider, J. Pruessner, H. Tsao, D. E. Fisher and C. L. Evans, Sci. Rep., 2016, 6, 37986.
222. D. A. Orringer, B. Pandian, Y. S. Niknafs, T. C. Hollon, J. Boyle, S. Lewis, M. Garrard, S. L. Hervey-Jumper, H. J. L. Garton, C. O. Maher, J. A. Heth, O. Sagher, D. A. Wilkinson, M. Snuderl, S. Venneti, S. H. Ramkissoon, K. A. McFadden, A. Fisher-Hubbard, A. P. Lieberman, T. D. Johnson, X. S. Xie, J. K. Trautman, C. W. Freudiger and S. Camelo-Piragua, Nat. Biomed. Eng., 2017, 1, 0027.
223. Y. Otsuka, S. Satoh, M. Kyogaku, H. Hashimoto, K. Itoh and Y. Ozeki, SPIE BiOS, 2014, pp. 89470C–89470C.
224. Y. Ozeki, F. Dake, S. Kajiyama, K. Fukui and K. Itoh, Opt. Express, 2009, 17, 3651–3658.
225. B. G. Saar, C. W. Freudiger, J. Reichman, C. M. Stanley, G. R. Holtom and X. S. Xie, Science, 2010, 330, 1368–1370.
226. C. W. Freudiger, R. Pfannl, D. A. Orringer, B. G. Saar, M. Ji, Q. Zeng, L. Ottoboni, W. Ying, C. Waeber, J. R. Sims, P. L. De Jager, O. Sagher, M. A. Philbert, X. Xu, S. Kesari, X. S. Xie and G. S. Young, Lab. Invest., 2012, 92, 1492–1502.
227. Y. Ozeki, W. Umemura, Y. Otsuka, S. Satoh, H. Hashimoto, K. Sumimura, N. Nishizawa, K. Fukui and K. Itoh, Nat. Photonics, 2012, 6, 845–851.
228. P. Berto, E. R. Andresen and H. Rigneault, Phys. Rev. Lett., 2014, 112, 053905.
229. D. Fu, W. Yang and X. S. Xie, J. Am. Chem. Soc., 2017, 139, 583–586.
230. C. W. Freudiger, W. Yang, G. R. Holtom, N. Peyghambarian, X. S. Xie and K. Q. Kieu, Nat. Photonics, 2014, 8, 153–159.
231. P. Nandakumar, A. Kovalev and A. Volkmer, New J. Phys., 2009, 11, 033026.
232. C. W. Freudiger, W. Min, B. G. Saar, S. Lu, G. R. Holtom, C. W. He, J. C. Tsai, J. X. Kang and X. S. Xie, Science, 2008, 322, 1857–1861.
233. D. Fu, G. Holtom, C. Freudiger, X. Zhang and X. S. Xie, J. Phys. Chem. B, 2013, 117, 4634–4640.
234. M. S. Alshaykh, C.-S. S. Liao, O. E. Sandoval, G. Gitzinger, N. Forget, D. E. Leaird, J.-X. X. Cheng and A. M. Weiner, Opt. Lett., 2017, 42, 1548–1551.
235. C.-S. S. Liao, M. N. Slipchenko, P. Wang, J. Li, S.-Y. Y. Lee, R. A. Oglesbee and J.-X. X. Cheng, Light: Sci. Appl., 2015, 4, e265.
236. R. Mittal, M. Balu, T. Krasieva, E. O. Potma, L. Elkeeb, C. B. Zachary and P. WilderSmith, Lasers Surg. Med., 2013, 45, 496–502.
237. F.-K. K. Lu, D. Calligaris, O. I. Olubiyi, I. Norton, W. Yang, S. Santagata, X. S. Xie, A. J. Golby and N. Y. R. Agar, Cancer Res., 2016, 3451–3462.
238. M. Ji, S. Lewis, S. Camelo-Piragua, S. H. Ramkissoon, M. Snuderl, S. Venneti, A. Fisher-Hubbard, M. Garrard, D. Fu, A. C. Wang, J. A. Heth, C. O. Maher, N. Sanai, T. D. Johnson, C. W. Freudiger, O. Sagher, X. S. Xie and D. A. Orringer, Sci. Transl. Med., 2015, 7, 309ra163.
239. S. Cui, P. Wang and S. Yue, Proc. SPIE, 2017, 100690M-1.
240. M. Egawa, K. Tokunaga, J. Hosoi, S. Iwanaga and Y. Ozeki, J. Biomed. Opt., 2016, 21, 86017.
241. R. Galli, O. Uckermann, A. Temme, E. Leipnitz, M. Meinhardt, E. Koch, G. Schackert, G. Steiner and M. Kirsch, J. Biophotonics, 2016, 404–414.
242. S. Yue, J. Li, S.-Y. Y. Lee, H. Lee, T. Shao, B. Song, L. Cheng, T. Masterson, X. Liu, T. Ratliff and J.-X. X. Cheng, Cell Metab., 2014, 19, 393–406.
243. X. Zhang, M. B. J. Roeffaers, S. Basu, J. R. Daniele, D. Fu, C. W. Freudiger, G. R. Holtom and X. S. Xie, ChemPhysChem, 2012, 13, 1054–1059.
244. Z. Wang, W. Zheng, J. Lin and Z. Huang, Proc. SPIE, 2016, 9712, 97120G.
245. A. L. Tam, H. J. Lim, I. I. Wistuba, A. Tamrazi, M. D. Kuo, E. Ziv, S. Wong, A. J. Shih, R. J. Webster, G. S. Fischer, S. Nagrath, S. E. Davis, S. B. White and K. Ahrar, J. Vasc. Interv. Radiol., 2015, 27, 8–19.
246. C. J. Frank, R. L. McCreery and D. C. Redd, Anal. Chem., 1995, 67, 777–783.
247. M. Balu, G. Liu, Z. Chen, B. J. Tromberg and E. O. Potma, Opt. Express, 2010, 18, 2380–2388.
248. B. G. Saar, R. S. Johnston, C. W. Freudiger, X. S. Xie and E. J. Seibel, Opt. Lett., 2011, 36, 2396–2398.
249. A. A. Hammoudi, K. K. Wong, L. Gao, P. Luo, S. T. C. Wong, Y. Yang and Z. Wang, Opt. Express, 2011, 19, 7960–7970.
250. S. Heuke, N. Vogler, T. Meyer, D. Akimov, F. Kluschke, H.-J. J. Röwert-Huber, J. Lademann, B. Dietzek and J. Popp, Healthcare, 2013, 1, 64–83.
251. C.-S. Liao and J.-X. Cheng, Sci. Adv., 2015, E1500738.
252. B. Spivack, M. M. Khanna, L. Tafra, G. Juillard and A. E. Giuliano, Arch. Surg., 1994, 129, 952–957.
253. M. Lacroix, D. Abi-Said, D. R. Fourney, Z. L. Gokaslan, W. Shi, F. DeMonte, F. F. Lang, I. E. McCutcheon, S. J. Hassenbusch, E. Holland, K. Hess, C. Michael, D. Miller and R. Sawaya, J. Neurosurg., 2001, 95, 190–198.
254. M. Jermyn, K. Mok, J. Mercier, J. Desroches, J. Pichette, K. Saint-Arnaud, L. Bernstein, M.-C. C. Guiot, K. Petrecca and F. Leblond, Sci. Transl. Med., 2015, 7, 274ra19.
255. K. Kong, F. Zaabar, E. Rakha, I. Ellis, A. Koloydenko and I. Notingher, Phys. Med. Biol., 2014, 59, 6141.
256. M. G. Shim and B. C. Wilson, Photochem. Photobiol., 1996, 63, 662–671.
257. R. Galli, O. Uckermann, E. Koch, G. Schackert, M. Kirsch and G. Steiner, J. Biomed. Opt., 2013, 19, 071402.
258. K. Klein, A. Gigler, T. Aschenbrenner, R. Monetti, W. Bunk, F. Jamitzky, G. Morfill, R. Stark and J. Schlegel, Biophys. J., 2012, 102, 360–368.
259. S. Gröschel, M. Bommer, B. Hutter, J. Budczies, D. Bonekamp, C. Heining, P. Horak, M. Fröhlich, S. Uhrig, D. Hübschmann, C. Geörg, D. Richter, N. Pfarr, K. Pfütze, S. Wolf, P. Schirmacher, D. Jäger, C. von Kalle, B. Brors, H. Glimm, W. Weichert, A. Stenzinger and S. Fröhling, Cold Spring Harbor Mol. Case Stud., 2016, 2, a001180.
260. G. Mountzios, M.-A. A. Dimopoulos, J.-C. C. Soria, D. Sanoudou and C. A. Papadimitriou, Crit. Rev. Oncol./Hematol., 2010, 75, 94–109.
261. N. Navin, A. Krasnitz, L. Rodgers, K. Cook, J. Meth, J. Kendall, M. Riggs, Y. Eberling, J. Troge and V. Grubor, Genome Res., 2010, 20, 68–80.
262. C. Curtis, S. P. Shah, S.-F. F. Chin, G. Turashvili, O. M. Rueda, M. J. Dunning, D. Speed, A. G. Lynch, S. Samarajiwa, Y. Yuan, S. Gräf, G. Ha, G. Haffari, A. Bashashati, R. Russell, S. McKinney, M. Group, A. Langerød, A. Green, E. Provenzano, G. Wishart, S. Pinder, P. Watson, F. Markowetz, L. Murphy, I. Ellis, A. Purushotham, A.-L. L. Børresen-Dale, J. D. Brenton, S. Tavaré, C. Caldas and S. Aparicio, Nature, 2012, 486, 346.
263. R. Natrajan, H. Sailem, F. K. Mardakheh, M. A. Garcia, C. J. Tape, M. Dowsett, C. Bakal and Y. Yuan, PLoS Med., 2016, 13, e1001961.
264. T. Ideguchi, S. Holzner, B. Bernhardt, G. Guelachvili, N. Picque and T. W. Hansch, Nature, 2013, 502, 355–358.
265. K. Hashimoto, M. Takahashi, T. Ideguchi and K. Goda, Sci. Rep., 2016, 6, 21036.
266. E. Ploetz, S. Laimgruber, S. Berner, W. Zinth and P. Gilch, Appl. Phys. B: Lasers Opt., 2007, 87, 389–393.
267. W. Rock, M. Bonn and S. H. Parekh, Opt. Express, 2013, 21, 15113.
268. K. Shi, H. Li, Q. Xu, D. Psaltis and Z. Liu, Phys. Rev. Lett., 2010, 104, 093902.
269. C. Heinrich, A. Hofer, A. Ritsch, C. Ciardi, S. Bernet and M. Ritsch-Marte, Opt. Express, 2008, 16, 2699–2708.
270. I. Toytman, K. Cohn, T. Smith, D. Simanovskii and D. Palanker, Opt. Lett., 2007, 32, 1941–1943.
271. M. Lei, M. Winterhalder, R. Selm and A. Zumbusch, J. Biomed. Opt., 2011, 16, 021102.
272. P. Berto, D. Gachet, P. Bon, S. Monneret and H. Rigneault, Phys. Rev. Lett., 2012, 109, 093902.
273. M. N. Slipchenko, R. A. Oglesbee, D. Zhang, W. Wu and J.-X. X. Cheng, J. Biophotonics, 2012, 801–807.

---

Footnote

† Official contribution of the National Institute of Standards and Technology; not subject to copyright in the United States.

---