Comparative evaluation of CAM methods for enhancing explainability in veterinary radiography.

Area of Science:

• Veterinary Radiology
• Explainable Artificial Intelligence (XAI)
• Deep Learning Interpretability

Background:

• Explainable Artificial Intelligence (XAI) methods, particularly Class Activation Mapping (CAM), are crucial for understanding deep learning model transparency in medical imaging.
• Systematic evaluations of CAM methods in veterinary radiography are limited, hindering their clinical adoption.
• This research addresses the gap by comparing multiple CAM techniques for visual interpretability in animal X-rays.

Purpose of the Study:

• To conduct a comparative analysis of eleven Class Activation Mapping (CAM) methods for explainable AI in veterinary radiography.
• To evaluate the interpretability and clinical relevance of heatmaps generated by different CAM techniques on canine and feline X-ray images.
• To assess the impact of these methods on veterinarians' diagnostic confidence.

Main Methods:

• Eleven CAM methods, including GradCAM, XGradCAM, ScoreCAM, and EigenCAM, were applied to a dataset of 7362 canine and feline X-ray images.
• A ResNet18 model was utilized, selected for its performance on the specific veterinary dataset.
• Quantitative and qualitative assessments were performed to evaluate heatmap interpretability and clinical relevance.

Main Results:

• EigenGradCAM, EigenCAM, and GradCAM++ demonstrated higher mean scores in interpretability compared to methods like FullGrad and XGradCAM.
• Despite performance variations, no single CAM method consistently improved veterinarians' diagnostic confidence across all cases.
• While some CAM techniques offered better visual cues for specific pathologies, overall explainability and diagnostic support were limited.

Conclusions:

• Current CAM methods provide varying degrees of visual interpretability in veterinary radiography but do not substantially enhance diagnostic confidence.
• Further research is needed to develop more effective XAI techniques tailored for veterinary diagnostic imaging.
• The study highlights the need for robust evaluation frameworks for XAI in specialized medical fields.