From Description to Diagnostics: Assessing AI's Capabilities in Forensic Gunshot Wound Classification

Area of Science:

• Forensic Science
• Medical Imaging
• Artificial Intelligence

Background:

• AI integration in forensic science is growing, but its use in diagnosing firearm injuries is limited.
• This study assesses ChatGPT-4's ability to differentiate gunshot wound entrance from exit wounds.

Purpose of the Study:

• To evaluate ChatGPT-4's diagnostic capabilities for firearm injuries.
• To determine the potential and limitations of AI in forensic pathology.
• To assess AI's applicability in distinguishing entrance from exit wounds.

Main Methods:

• ChatGPT-4 was tested on three datasets: external firearm injury images, intact skin images, and real-case firearm injury images.
• AI performance was evaluated before and after machine learning (ML) training.
• Classification accuracy was assessed using descriptive and inferential statistics.

Main Results:

• ChatGPT-4 significantly improved entrance wound identification after ML training, enhancing morphological descriptions.
• AI performance in classifying exit wounds remained limited, consistent with forensic literature challenges.
• The AI achieved 95% accuracy in distinguishing intact skin from injuries, but lacked contextual data for misclassifications.

Conclusions:

• ChatGPT-4 shows potential as a supplementary forensic diagnostic tool due to its learning capacity.
• Specialized deep learning models are currently superior to ChatGPT-4 for forensic applications.
• Expert oversight is crucial to mitigate risks like overconfident misclassifications and hallucinations in AI use.