Automated Estimation of Sperm DNA Fragmentation from Sperm Videos Using 3D ResNet and TimeSformer

Abstract

Sperm DNA fragmentation (SDF) is one of the key indicators for assessing male infertility. However, the currently widespread TUNEL-FACS method (a combination of the TUNEL assay and flow cytometry) is invasive to sperm and requires expensive specialized equipment. In this study, we propose a deep learning-based model that non-invasively estimates SDF values from sperm motility videos captured with standard phase-contrast microscopy. To capture temporal variations such as sperm motion, each video is divided into eight-frame clips with a four-frame overlap. We evaluate two backbone architectures 3D ResNet and TimeSformer in combination with spatial pre-processing techniques that enhance sperm visibility while suppressing background noise. To aggregate clip-level predictions into a video-level SDF score, we analyze three aggregation strategies: mean, median, and best-clip aggregation. The results show that TimeSformer, when combined with spatial pre-processing and median aggregation, achieves the highest practical performance. Although best-clip aggregation demonstrates the highest accuracy, it relies on ground-truth labels to select the best-performing clip and is therefore not suitable for clinical application. These findings suggest that introducing intelligent clip-selection mechanisms could further improve prediction accuracy. Our proposed approach holds promise as a novel, stain-free, and non-invasive method for SDF assessment.