Dynamic Multi-scale Class Activation Mapping (DMs-CAM) for Enhanced Explainability in Diabetic Retinopathy Classification

Explainability is crucial for clinical adoption of deep learning in medical imaging. Existing Class Activation Mapping (CAM) methods, such as Grad-CAM and LayerCAM, either produce coarse localization or suffer from background noise interference. We propose Dynamic Multi-scale CAM (DMs-CAM), which integrates global semantic guidance into local gradient weighting via a lightweight dynamic weight allocation strategy, and employs an SE-based multi-scale fusion mechanism to adaptively combine activation maps across network depths. Without introducing extra network layers, DMs-CAM enhances focus on salient regions and preserves multi-scale details. We validate DMs-CAM on Diabetic Retinopathy (DR) task using the OIA-DDR dataset: pathological grading (6 classes, VGG16 backbone). Quantitative evaluation and qualitative visualization demonstrate that DMs-CAM outperforms Grad-CAM, Grad-CAM ++, and LayerCAM in localization accuracy, noise suppression, and multi-scale consistency, particularly improving detection of subtle lesions. DMs-CAM advances interpretability and trustworthiness of CNNs in ophthalmology.