Explainable Medical Image Analysis

This field studies how AI systems analyze medical images and how clinicians can inspect the evidence behind predictions, segmentations, and reports. Approaches include post-hoc saliency and perturbation methods, prototypes and concept bottlenecks, interpretable radiomic features, counterfactuals, and vision-language systems that link findings to clinical concepts or anatomical regions. Methods are tested across tasks such as chest X-ray diagnosis, retinal grading, tumor segmentation, and whole-slide pathology, with growing attention to external-site generalization and clinician assessment. A central challenge is that attractive explanations may not faithfully reflect what drives a model, while explanation stability, clinical usefulness, and predictive performance can vary across architectures, methods, and datasets.