Benchmarking Class Activation Map Methods for Explainable Brain Hemorrhage Classification on Hemorica Dataset

CAM methods can localize hemorrhages without explicit training: The study found that Class Activation Mapping techniques successfully identified brain hemorrhage locations with a Dice score of 0.57 and IoU of 0.40, even though the AI models were trained only to classify images rather than to segment or locate hemorrhages—demonstrating that explainability methods can extract clinically useful information beyond what models were directly taught.

Different CAM techniques perform better at different stages: The research benchmarked nine different CAM algorithms and found that performance varied significantly depending where in the neural network the method was applied, with stage 5 of EfficientNetV2S showing the strongest results—suggesting that one-size-fits-all approaches to AI explainability may be insufficient for medical imaging.

Quantitative evaluation of explainability is now possible: This study established the first quantitative benchmark comparing CAM methods for brain hemorrhage detection using standard metrics (Dice, IoU, pixel-wise overlap), moving explainability research beyond subjective visual inspection and enabling more rigorous validation of which explanation methods are most trustworthy for clinical use.

High-quality datasets enable better explainability research: The Hemorica dataset's unique combination of both classification labels and detailed segmentation masks allowed researchers to rigorously evaluate whether CAM explanations actually pointed to the right anatomical locations—highlighting that dataset design choices directly impact our ability to validate AI transparency tools.