A Multi-CAM Based Robust Interpretations Framework in Medical Image Analysis
Yeong-Eun Jeon, Minseo Hwangbo, Minyoung Heo, Ho-Jung Kim, Ga‐Hyun Son, Dong-Ok Won · 2025
Recently, methods for interpreting the decision of deep learning models have been actively used in medical imaging. Still, some studies provide interpretations based solely on a singular method, such as GradCAM. This approach has the limitation to reduce reliability by providing only fragmentary information about model predictions. In this study, we propose a framework that multi-interpretations of model results to improve the reliability of deep learning based diagnostic assistance systems. We combined several CAM based methods and deep learning model and analyzed the differences in interpretation based on the CAM methods. The results show that the same image has different interpretations depending on the CAM, which means that multiple interpretations through a combination of appropriate CAM and deep learning models are essential. Consequently, the proposed framework can improve diagnostic reliability and increase the usability of AI-based diagnostic assistance systems by providing multiple interpretations of model predictions.