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

Area of Science:

• Forensic Science
• Artificial Intelligence
• Computer Vision

Background:

• Accurate classification of sharp force injuries is crucial in forensic pathology.
• Traditional methods rely on expert human analysis, which can be time-consuming and subjective.

Purpose of the Study:

• To develop and evaluate an AI-based deep learning method for automatic identification and classification of sharp injuries.
• To assess the accuracy and explainability of AI models in supporting forensic injury classification.

Main Methods:

• A dataset of 1161 sharp injury images (stab, chop, slash) was utilized.
• Three deep learning models (ResNet50, GoogLeNet, ShuffleNet-V2) were fine-tuned using transfer learning.
• AI model performance was quantitatively tested and compared against forensic pathologists using external data.

Main Results:

• The GoogLeNet model achieved 88.2% overall accuracy, with high accuracy for stab (98.4%) and chop wounds (96.7%).
• AI classification was significantly faster than human analysis, with comparable accuracy for stab and chop wounds.
• AI model explanations (saliency maps) aligned with forensic pathologists' observations.

Conclusions:

• AI models demonstrate practical utility in forensic injury classification, particularly for stab and chop wounds.
• The developed AI method enables accurate recognition and rapid differentiation of sharp force injuries.
• AI shows potential to augment forensic pathology expertise, improving efficiency and consistency.