MPCNet: multi-scale decomposition network with perception weaving and context-aware fusion for robust 2D medical image classification.

Objective.Accurate medical image classification is crucial for improving diagnostic efficiency and reliability. However, existing approaches face significant challenges in modeling global dependencies, balancing fine-grained and coarse-grained features, and integrating multi-scale representations, which limits their adaptability across diverse datasets and lesion types.Approach.To address these challenges, we propose a multi-scale decomposition network with perception weaving and context-aware fusion (MPCNet) for Robust 2D Medical Image Classification. The multi-scale decomposition block is designed to capture fine-grained features through hierarchical convolutions, enabling precise extraction of subtle lesion details and linear structures. The perception weaving block is developed to model coarse-grained features by combining a dual-pooling strategy with dynamic attention, thereby enhancing global perception and multi-scale integration. The context-aware fusion block adaptively integrates complementary fine- and coarse-grained features from two branches via spatial and channel attention mechanisms, dynamically adjusting fusion weights based on contextual relevance to substantially improve classification accuracy.Main results.MPCNet achieves outstanding performance across multiple medical image classification datasets: 87.54% accuracy on Ulcerative Colitis, 81.98% on Kvasir, 73.68% on COVID-19, 72.81% on PAD-UFES-20, and 96.57% on Fetal-Planes-DB, demonstrating high adaptability to diverse disease types and data characteristics. Comprehensive ablation and comparative experiments confirm the effectiveness of each module. Extensive experimental results show that the proposed method delivers state-of-the-art performance across multiple classification tasks.SignificanceThe experimental findings fully demonstrate the effectiveness and superior generalization capability of our MPCNet, offering a robust and unified solution for medical image analysis. The code is publicly available atour GitHub repository.