AI-Assisted Forensic Analysis of Hanging-Related Ligature Marks: A Pilot Study Using Convolutional Neural Networks.
Authors
Affiliations (3)
Affiliations (3)
- Department of Biomedical and Dental Sciences and Morphofunctional Imaging, University of Messina, Via Consolare Valeria, 1, 98125 Messina, Italy.
- Department of Engineering, University of Messina, Contrada Di Dio, Sant'Agata, 98166 Messina, Italy.
- State Forensic Science, Mykolas Romeris University Vilnius, 01101 Vilnius, Lithuania.
Abstract
<b>Background:</b> Artificial intelligence (AI) is increasingly applied in medical image analysis, although its application in forensic pathology remains limited. The assessment of ligature marks in hanging deaths is challenging and relies on forensic expertise. This pilot study evaluated a deep learning approach for morphological classification of hanging-related ligature marks. <b>Methods:</b> A Convolutional Neural Network (CNN) was trained on a dataset of 404 standardized JPEG images obtained from forensic medicine atlases and classified into hanging-related ligature marks and non-hanging lesions, including strangulation and post-mortem artefacts. Following internal validation, the model was tested on an independent set of forensic case images provided by forensic pathology experts from Messina, Italy, and Vilnius, Lithuania. Images were annotated and reviewed by a team of two forensic pathologists, with final labels assigned by consensus using morphological criteria. <b>Results:</b> The CNN demonstrated encouraging classification performance with an F1-score of 0.81 ± 0.04, distinguishing hanging-related ligature marks from morphologically similar lesions. The methodological framework and image standardization criteria for AI-assisted forensic analysis were also established. <b>Conclusions:</b> AI-based image analysis may support the evaluation of ligature marks during external examinations. Nevertheless, forensic diagnosis requires the integration of autopsy findings, physical examination, and circumstantial evidence. Larger datasets and multicenter protocols are needed to further assess reliability and applicability.