When LLMs Analyze Scars: From Images to Clinically-Meaningful Features

The authors propose a novel framework called ScaFE (Scar Feature Engineering) that utilizes large language models (LLMs) as feature engineers for medical image classification, specifically for pathological scar differentiation. By generating executable Python code based on established clinical criteria, the approach enhances data efficiency, preserves privacy by processing images locally, and improves interpretability of the features extracted. Experimental results indicate that ScaFE outperforms traditional end-to-end deep learning models in scenarios with limited annotated data, highlighting its potential for developing clinically transparent AI systems in healthcare.