Sperm detection remains a critical step in sexual assault investigations. While sperm-rich slides can be rapidly classified as positive, weakly positive or negative slides often require extensive manual microscopic examination. This process represents a major analytical bottleneck in forensic genetics laboratories.

At the Institut Génétique Nantes Atlantique (IGNA), several commercial artificial intelligence solutions were initially evaluated to automate sperm detection. Systems such as Calopix (Tribun) and Metafer (Metasystems) offer automated detection workflows based on whole-slide scanning or real-time motorized microscopy. Although these systems demonstrated promising capabilities, they also presented important limitations, including reliance on proprietary image libraries, limited transparency regarding the algorithms, and vendor-dependent retraining procedures.

To overcome these constraints, we decided to develop a fully laboratory-controlled solution, trained on our own data and adapted to our protocols.

This approach led to the development of Deep-Zoïde, an artificial intelligence platform coupled with the implementation of the Evident VS200 high-resolution whole-slide scanner. The VS200 enables multi-plane scanning with oil immersion, producing a consensus image with uniform sharpness across the entire slide and closely mimicking the multi-focus exploration performed during manual microscopy.

Whole-slide images in multi-gigabyte .tif format are analyzed using a patch-based inference pipeline based on 640 × 640 pixel tiles compatible with YOLO detection models. An optimized “grid-only” architecture computes the tiling geometry and performs dynamic in-memory cropping during GPU inference, avoiding the generation of intermediate patch files and reducing disk I/O operations.

The detection model was trained on a dataset of more than 100,000 images annotated by forensic experts. Final model performance metrics (Precision : 0.93 ; Recall : 0.90 ; mAP50 : 0957 ; mAP50-95 : 0.799)  demonstrate excellent detection performance, with high precision, strong recall, and very good mAP values.

A validation study on 100 production slides was conducted by comparing Deep-Zoïde results with expert microscopic examination and demonstrates that Deep-Zoïde can achieve performance comparable to, or even exceeding, human expertise.

Although such performance suggests that a fully automated workflow could be possible, Deep-Zoïde is currently implemented as a decision-support tool. Experts can review detections, correct errors, and feed these corrections back into the dataset, enabling continuous model fine-tuning through an active learning loop.

Deep-Zoïde therefore provides a transparent, scalable, and fully controlled framework for integrating artificial intelligence into forensic sperm detection workflows.

Authors

• Catherine Dehainault (Institut Génétique Nantes Atlantique, France)
• Yann Chovory (IA group Carso, France)
• Maxime Eynard (Institut Génétique Nantes Atlantique, France)
• Sandra Rabu (Institut Génétique Nantes Atlantique, France)
• Patricia Michel (Institut Génétique Nantes Atlantique, France)
• Muriel Garnier (Institut Génétique Nantes Atlantique, France)
• Marie-Gaëlle Le Pajolec (Institut Génétique Nantes Atlantique, France)
• Soizic Le Guiner (Institut Génétique Nantes Atlantique, France)