Enhanced semen stain detection through the synergistic integration of a prototype artificial intelligence (AI) model and polilight technique

Abstract

The increasing caseload of sexual offenses in Thailand places a significant strain on forensic science resources, demanding more efficient methods for evidence analysis. Preliminary screening of semen stains using Polilight is standard practice, but differentiating semen fluorescence from other visually similar biological fluids remains a challenge, often leading to time-consuming and labor-intensive confirmatory tests. This research developed and evaluated an artificial intelligence (AI) system to enhance the accuracy of semen stain detection. Two Roboflow 3.0 Object Detection (Accurate) models, based on a YOLOv8-compatible architecture, were developed. The models were trained on a dataset of 1597 images comprising semen (n = 719), urine (n = 301), and milk (n = 577) stains under two lighting conditions: "under direct light" and "outside the direct light radius". One model was trained with data augmentation and the other without. The performance of both models was validated and compared against the analysis of five experienced forensic experts. The model trained with data augmentation demonstrated significantly superior performance. Under direct light, it achieved an overall accuracy of 75.00 %, outperforming the 64.50% accuracy of the forensic experts. In challenging lighting conditions (outside the direct light radius), the experts exhibited a highly conservative strategy, achieving exceptional specificity (94.00 %) but very low sensitivity (45.00 %). In contrast, the AI models maintained a more balanced and consistent performance. Qualitative analysis confirmed the AI's reliability in identifying semen but also highlighted challenges in differentiating it from urine and milk, a limitation also observed in human experts. An AI model enhanced with data augmentation is a robust, consistent, and effective tool for the preliminary screening of semen stains, capable of exceeding the accuracy of human experts in ideal conditions. The findings support a synergistic human-AI workflow, where the AI serves as a standardized primary screening tool to increase efficiency and reduce workload, followed by expert verification for ambiguous cases. This approach holds the potential to accelerate the investigative process and strengthen the reliability of forensic conclusions.