Multi-scale dual-path cascaded attention network for skin lesion classification

Abstract

Skin cancer represents a major global health concern, where early detection is critical for improving patient outcomes. Despite considerable promise in dermoscopic image analysis, existing deep learning methods share three persistent limitations: single-stream architectures cannot simultaneously capture global lesion morphology and local fine-grained details; fixed-receptive-field convolutions fail to handle wide intra-class scale variation; and parallel attention modules lack a structured hierarchical refinement logic. To address these issues, we propose the multi-scale dual-path cascaded attention network (MDCAN). A dual-path architecture explicitly decouples global semantic and local detail representations, while a cascaded attention module (MSCA) enforces a sequential ‘channel → spatial → scale’ refinement pipeline within each path. Depthwise separable and dilated convolutions are incorporated to capture information across different receptive fields while limiting additional computational overhead. Evaluated on the ISIC2017, ISIC2019, and HAM10000 datasets over five independent runs, MDCAN achieved mean accuracies of 89.25 ± 0.22%, 91.82 ± 0.32%, and 93.81 ± 0.17%, respectively. With 17.49 M parameters, MDCAN demonstrates competitive classification performance and a favorable trade-off between classification accuracy and model complexity.