Gaussian clustering and quantification of the sperm chromatin dispersion test using convolutional neural networks

Abstract

Sperm DNA fragmentation is a sign of sperm nuclear damage. The sperm chromatin dispersion (SCD) test is a reliable and economical method for the evaluation of DNA fragmentation. However, the cut-off value for differentiation of DNA fragmented sperms is fixed at 1/3 with limited statistical justification, making the SCD test a semi-quantitative method that gives user-dependent results. We construct a collection of deep neural networks to automate the evaluation of bright-field images for SCD tests. The model can detect valid sperm nuclei and their locations from the input images captured with a 20× objective and predict the geometric parameters of the halo ring. We construct an annotated dataset consisting of N = 3120 images. The ResNet 18 based network reaches an average precision (AP₅₀) of 91.3%, a true positive rate of 96.67%, and a true negative rate of 96.72%. The distribution of relative halo radii is fit to the multi-peak Gaussian function (p > 0.99). DNA fragmentation is regarded as those with a relative halo radius 1.6 standard deviations smaller than the mean of a normal cluster. In conclusion, we have established a deep neural network based model for the automation and quantification of the SCD test that is ready for clinical application. The DNA fragmentation index is determined using Gaussian clustering, reflecting the natural distribution of halo geometry and is more tolerable to disturbances and sample conditions, which we believe will greatly improve the clinical significance of the SCD test.