Deep learning and firearm wound classification: a pilot study

Le résumé fourni par la source

Introduction: The study on the use of deep learning in pattern recognition of gunshot wounds (GSW) represents a novelty in the field of forensic pathology. Although artificial intelligence (AI) has already revolutionized many medical specialties, applications in forensic medicine are still limited. Nevertheless, AI-based tools could be of great use in a discipline that relies heavily on visual analysis. Scant scientific evidence from recent experimental studies has demonstrated the AI potential in predicting shooting distance based on GSW pictures. Previous approaches achieved a classification accuracy of 98%; however, further studies are needed to evaluate its applicability in forensic practice. The aim of this project is to further explore the application of deep learning techniques for the classification of GSWs. Materials and methods: The study, conducted at the University of Catania, employed the free software Lobe AI. It was necessary to proceed through 4 consecutive phases: training, model validation, testing, and data analysis. Four study categories were then identified: GSWs, entrance/exit wounds, wounds based on the pathological range of fire, and wounds based on the type of weapon ammunition. For the software training, educational images were extracted from a forensic atlas specializing in GSWs. For the testing phase, photos from the Catania forensic case history and photos of healthy skin as a control were selected. For each study category, two phases were adopted, characterized by an increasing number of files used for training. Finally, data were recorded and subjected to statistical analysis based on the following parameters: accuracy, precision, recall, F1-score, specificity. Results: Encouraging data have been observed, especially when compared with available scientific evidence. Numerous parameters were shown to be above the threshold attributed to the human limit, and in some fields, the highest predictive values to date were recorded. Discussion: Among the intrinsic limitations of the study, the limited data available emerged, especially for the algorithm training phase; moreover, it will be vital to proceed with the development of a specific software pre-trained on the basis of natural elements and secondarily directed to the recognition of the typical forensic characteristics. Among the strengths, however, it is possible to include the number of categories analyzed and the unprecedented use of an "intact skin" control category. In conclusion, the extension of the project to various research centers and the increase in the size of the sample adopted will allow the elimination of the emerged biases. This pilot study lays the groundwork for future multicenter research aimed at developing a robust and generalizable forensic AI model.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.