Spitz: a YOLO-based deep learning system for automated spermatozoa detection in forensic genetic analysis

In sexual assault investigations, microscopic spermatozoa identification is a critical yet subjective step in the forensic genetic workflow. Manual examination is time-consuming, prone to inter-examiner variability, and susceptible to false negatives. This study presents Spitz, a deep learning-based system for automated sperm cell detection in forensic microscopy samples and evaluates its performance against manual examination through blind comparative study. A YOLOv10-L model was trained on 399 microscopic image patches from real-world forensic cases stained with Christmas Tree stain. The model was integrated into a web interface enabling batch inference, spatial slide reconstruction, and coordinate-based expert verification. A blind comparative study was conducted with three forensic experts re-examining 30 sexual assault samples using both manual and AI-assisted methods, recording spermatozoa counts, classification (Not present, Rare, Occasional), and examination time. Statistical analysis included Wilcoxon signed-rank tests, Fleiss' Kappa, and coefficient of variation (CV). The model achieved 94% precision, 72.3% recall, and mAP50 of 0.842. AI-assisted detection found significantly more spermatozoa than manual examination, with lower count variability. Inter-rater agreement improved (Fleiss' Kappa: 0.858 vs. 0.698; agreement: 86.7% vs. 73.3%), with reduced classification variability across categories (mean CV: 80.1% vs. 93.5%). Examination time was shorter with AI (6.39 ± 0.59 min vs. 6.93 ± 0.70 min), though not statistically significant (p = 0.2621). Spitz significantly surpasses manual examination in detection sensitivity and inter-examiner reproducibility while reducing examiner light exposure. By integrating detection, coordinating tracking, and microscope navigation, it bridges automated analysis and downstream forensic genetic workflows.