A machine learning approach for automated injuries classification on postmortem images.

The application of Artificial Intelligence (AI) in forensic science offers new opportunities for the automated detection of injuries in postmortem analysis. This study focuses on the semantic segmentation of two significant types of injuries-bruises and abrasions. A dataset of postmortem injury images was collected, followed by the development of appropriate data preprocessing and annotation techniques to train and evaluate AI models. Three deep learning architectures-U-Net, FPN, and LinkNe-were implemented, using EfficientNetB3 and ResNet50 as backbone networks. An optimisation strategy was employed to enhance detection performance by incorporating a custom loss function alongside a combination of image transformation and class balancing techniques. Experimental results demonstrated high sensitivity (92.7 %) and specificity (98.9 %) for the best-performing model. These findings highlight the potential of AI-driven methods for automated and objective analysis of postmortem images in injury detection, laying the foundation for further research.