Abstract
Robust classification of medical images is crucial for reliable automated diagnosis, yet remains challenging due to heterogeneous lesion appearances and imaging inconsistencies. We introduce DWAM-MSFINET (Dual Window Adaptation and Multi-Scale Feature Integration Network), a novel deep neural architecture designed to address these complexities through a dual-pathway integration of attention and resolution-aware representation learning. Specifically, the Multi-Scale Feature Integration (MSFI) module hierarchically aggregates semantic cues across spatial resolutions, enhancing the network's capacity to identify both fine-grained and coarse pathological patterns. Complementarily, the Dual Weighted Attention Mechanism (DWAM) adaptively modulates feature responses in both spatial and channel dimensions, enabling selective focus on clinically salient structures. This unified framework synergizes localized sensitivity with global semantic coherence, effectively mitigating intra-class variability and improving diagnostic generalization. DWAM-MSFINET achieved 78.6% Top-1 accuracy on the standalone Messidor dataset, demonstrating robustness against domain shift. DWAM-MSFINET surpasses state-of-the-art CNN and Transformer-based models, achieving a Top-1 accuracy of 82.59%, outperforming ResNet50 (81.68%) and Swin Transformer (80.26%), while inference latency is 16.0 ms per image (not seconds) when processing batches of 16 images on NVIDIA RTX 3090, equivalent to 62.5 images per second. These results validate the efficacy of our approach for scalable, real-time medical image analysis in clinical workflows. We have released our code and datasets at: https://github.com/eleen7/data.