Explainable AI for sharp injury identification using transfer learning with pre-trained deep neural networks

Objective

To investigate an AI-based method for automatically identifying and classifying sharp injuries using deep learning models, evaluate its effectiveness (e.g., accuracy and explainability), and support forensic injury classification.

Methods

A dataset comprising 1161 photos was collected, including stab wounds (723), chop wounds (314), and slash wounds (124) from homicide cases. After preprocessing and weighted random sampling, the processed dataset was divided into training and validation sets in an 8:2 ratio. Additionally, 212 images from new cases representing the three types of wounds were collected as an external dataset for a human vs. AI test. Specifically, three classification networks pre-trained on natural images—ResNet50, GoogLeNet, and ShuffleNet-V2—were fine-tuned via transfer learning on the training set. The models were then quantitatively tested in terms of precision, recall, F1 score, and reading time. The test results of AI models were compared with forensic pathologists using the external test data. Moreover, we analyzed the image explanation factors captured by these models according to the saliency maps produced by the class activation mapping techniques.

Results

All three models successfully classified three types of wounds. Among these, the GoogLeNet network model demonstrated an overall classification accuracy (total) and recall rate of 88.2 %. The model achieved its highest classification accuracy of 98.4 % for stab wounds, followed by 96.7 % for chop wounds and 30.0 % for slash wounds, the lowest among them. Classification accuracy is positively correlated with sample size. The model achieved a maximum precision rate of 88.4 % and a F1 score of 0.860, with a classification time of 0.04 seconds per image. A comparison with forensic pathologists revealed that the model's classification time was shorter, while its accuracy of stab and chop wounds was comparable to that of senior forensic pathologists, but the accuracy for slash wound was lower than that of junior forensic pathologists. The image explanation factors captured by AI models align closely with the characteristic wound positions identified by forensic pathologists.

Conclusion

The AI model effectively identifies the image characteristics of stab and chop wounds, enabling accurate recognition and rapid differentiation. The AI classification performance for stab and chop was comparable to that of senior forensic pathologists, implying the model’s practical utility.