Deep learning for subtypes identification of pure seminoma of the testis

Abstract

Pathological evaluation of each tumor sample is the most crucial process in the clinical diagnosis workflow. Deep learning is a powerful approach that is widely used to increase accuracy and to simplify the diagnosis process. Previously we discovered clinically relevant subtypes (1 and 2) of pure seminoma, which is the most common histological type of testicular germ cell tumors (TGCTs). Here we developed deep learning decision making tool for identification of seminoma subtypes using histopathological slides. We used all available slides for pure seminoma samples from The Cancer Genome Atlas (TCGA). Seminoma regions of interest (ROIs) were annotated by a genitourinary pathologist. Verified ROIs were split into tiles 300×300 pixels, which were used for model training. The model achieved the highest accuracy at the validation step with 0.933 with 0.92 area under the ROC curve.