End-to-End Spermatozoid Detection in Cytology WSI for Forensic Pathology Workflow

Rutger RH Fick, Melanie Lubrano, Damien Quignon, Didier Empis, Fabrice Kabile, Saima Ben Hadj

Keywords: Spermatozoid screening, Whole Slide Images (WSI), Deep learning, Forensic laboratories, Object detection

TL;DR: This study introduces a deep-learning algorithm to boost sensitivity and throughput in forensic spermatozoid screening using cytology Whole Slide Imaging, effectively detecting missed spermatozoids.

Abstract: This study aimed to improve the sensitivity and throughput of spermatozoid screening for identifying rape suspects through DNA profiling, based on microscope cytology Whole Slide Imaging (WSI). To this end, we implemented a WSI-based deep-learning algorithm consisting of a detector/classification ensemble, achieving a mean 3-fold cross-validation F1 score of 0.87 [0.87-0.88] on a dataset of 188 retrospective single-center cytology WSI. Applied to slide label-only annotated test set (positive, negative, and doubtful), we show that our ensemble model is capable of screening slide label groups with excellent sensitivity to even find missed spermatozoids in negative-labeled slides. We hope our approach will be of value for routine forensic spermatozoid screening.