Explainable Deep Learning for Forensic Wound Analysis: A Survey of Injury Classification, Gunshot Wound Interpretation, and Toward Court-Admissible AI

Zakaria Ahmed, Department of Computer Science, Faculty of Computer and Informatics, Zagazig University, 44519, EgyptFollow
Zeshan Aslam Khan, International Graduate School of Artificial Intelligence, National Yunlin University of Science and Technology, 123 University Road, Section 3, Douliou, Yunlin 64002, Taiwan, R.O.CFollow

ORCID

Zakaria Ahmed: https://orcid.org/0009-0002-9654-1997

Zeshan Aslam Khan: https://orcid.org/0000-0003-3902-5814

Keywords

Explainable deep learning, Forensic wound analysis, Gunshot wound classification, Injury classification, Court-admissible AI, Explainable AI

Article Type

Original Article

Abstract

Wound analyses conducted in forensics continue to be based on visual interpretation of pathologists and therefore exhibit variations among observers, especially when dealing with unusual injuries. Recent research using deep learning algorithms has proven that automated identification of wounds and interpreting gunshot wounds has achieved success with high levels of accuracy, ranging between 88% on large human datasets in distinguishing exit/entrance wounds and 94% to 98% in shooting distance and shotgun patterns. Nonetheless, most of these models still operate as black boxes. In this systematic review of deep learning methods in forensic wound assessment, there is emphasis placed on three related topics: general injury classification, gunshot wound assessment, and the creation of court-ready AI systems. Using a PRISMA-based selection process of seven relevant primary papers, this review shows how although some AI approaches such as ResNet-152 and ConvNeXt Tiny have shown performance comparable to experts when classifying morphologically distinct wounds, methods for increasing model explainability (Grad-CAM, SHAP, LIME) are virtually nonexistent in the existing literature. Additionally, external validation is often not performed or done in very small sample sizes from one institution or even an animal study and no system in this review met standards for validity and reliability in terms of legal admissibility, including the Daubert standard and Federal Rule of Evidence 707. The review points out three crucial shortcomings in current AI approaches: explainability, validation, and legal admissibility, and calls for further approaches to have these factors in place.

How to Cite

Ahmed, Zakaria and Khan, Zeshan Aslam (2026) "Explainable Deep Learning for Forensic Wound Analysis: A Survey of Injury Classification, Gunshot Wound Interpretation, and Toward Court-Admissible AI," Sustainable Machine Intelligence Journal: Vol. 14: Iss. 3, Article 3.
DOI: https://doi.org/10.63689/3005-3617.1093