Deep Learning–Based Image Analysis Model for Classification and Quantification of Multiple Histopathological Findings in Rat Testis and Epididymis

Abstract

Supervised deep learning–based image analysis models using whole slide images (WSIs) have been reported to be effective for detecting simple histopathological findings in laboratory animals. However, there are no models that simultaneously detect multiple types of abnormal testicular findings on hematoxylin and eosin (H&E)-stained specimens in rats. In this study, we developed a model that can detect, classify, and quantify major 7 testicular toxicity findings in rats (degeneration of germ cell, tubular atrophy, tubular dilatation, vacuolation of Sertoli cell, multinucleated giant cell in the testis, and decreased sperm and cell debris in the epididymis) in addition to classify spermatogenic stages on H&E-stained WSIs. For training the model, we used WSIs of the testis and epididymis of rats that were administered various compounds in toxicity studies, and we developed it using supervised deep learning algorithms and WSI data sets. Detection accuracy of the spermatogenic stage classification and the 7 findings generated by the model was compared with histopathological diagnoses made by board-certified pathologists and confirmed to be high performance. Therefore, this model is a useful tool to support histopathological evaluation, especially in initial screening in rat toxicity studies and is expected to improve work efficiency and prevent errors due to oversights.