Multiclass Skin Lesion Classification Using a Hybrid Symmetry-Aware and Soft Attention (HSASA) Model

• Gururaj Nase,
• Arun Arthanareeswaran,

Vol. 19, No. 6, June 30, 2025
10.3837/tiis.2025.06.002, Download Paper (Free):

• Symmetry-Aware Cross-Attention (SACA)
• Skin Lesion Classification
• Attention Mechanism
• Melanoma
• Convolutional-neural-networks (CNNs)
• Skin Cancer (SC)

Abstract

Skin cancer is one of the most dangerous diseases worldwide. Correct classification of skin lesions (SLs) at an early stage can assist clinical decision-making by providing an accurate diagnosis, significantly improving the probability of a cure before the cancer spreads. However, automatic SC classification poses challenges due to imbalanced and limited datasets, along with issues related to model robustness and cross-domain adaptability. In this paper, we propose a novel model that utilizes a unique attention mechanism to enhance diagnostic precision. The model isolates and emphasizes spatial features and lesion symmetry, allowing for targeted analysis of class-specific differences, such as symmetry, texture, and color uniformity. To further capture the varying lesion boundaries across classes, we incorporate a soft attention mechanism that extracts detailed boundary information, improving classification by focusing on the most relevant regions in each image. This approach also enhances model interpretability, as the soft attention maps highlight the input areas influencing the model’s decisions, fostering greater trust among medical professionals. Additionally, we introduce a Symmetry-Aware Cross-Attention (SACA) module, which analyzes the relationship amongst an image and its symmetrical counterpart to better capture lesion characteristics. Validation on the HAM10000 dataset demonstrates that our model surpasses several state-of-the-art methods in accuracy, suggesting it as a reliable, interpretable tool for early SC diagnosis.

Statistics

Cite this article