Skin Disease Classification Algorithm Based on Multi-Scale Feature Decoupling and Boundary-Aware Diagnosis for Complex Lesion Morphology Recognition

Abstract

Skin lesion images often present challenges due to complex color distributions, subtle texture variations, irregular edge morphology, and indistinct visual differences between different disease categories, posing a challenge to the stable recognition of automated classification models. To address the insufficient representation of complex lesion morphologies by single global features and the weakening of local diagnostic cues, this paper proposes a multi-scale feature decoupling and boundary-aware diagnostic algorithm for skin disease classification. This method first obtains lesion representations at different scales through hierarchical feature extraction, enabling shallow texture details and deep semantic information to jointly participate in classification modeling. Subsequently, a feature decoupling mechanism is constructed, dividing the mixed features into morphological, texture, and boundary-related subspaces to reduce redundant interference between different visual cues and enhance the model's ability to express the heterogeneous structure of lesions. Based on this, an adaptive scale fusion strategy is introduced to dynamically integrate features at different levels, allowing the model to highlight more discriminative scale responses based on sample feature differences. Simultaneously, a boundary-aware calibration module uses lesion contour and transition region information to spatially enhance the fused features, thereby improving the model's ability to recognize blurred boundaries, local diffusion, and irregular morphologies. Comparative experiments based on publicly available skin lesion image datasets show that the proposed method achieves superior performance in terms of accuracy, precision, recall, and F1 score, validating the effectiveness of the proposed algorithm in complex skin lesion classification tasks.