MTA-Net: Multi-scale triplet attention-aware network for multiclass skin lesion classification

Multiclass classification of skin lesions plays a crucial role in computer-aided skin cancer diagnosis and still remains challenging due to the high similarity between different classes and large variations within the same classes. Additionally, the color, shape, and size variation of lesions raise significant concerns. Existing convolutional neural network (CNN)-based approaches have shown prominence over traditional feature engineering-based methods for learning high-level feature representations from dermoscopy images; however, they exhibit limited capabilities to capture subtle lesion variations. Although recent attention-driven CNN models have partially addressed the above issues, they struggle to learn fine-grained lesion information at various scales, hindering performance in multiclass skin lesion classification settings. To address these challenges, this paper proposes a multi-scale triplet attention-aware network (MTA-Net) for multiclass skin lesion classification. Specifically, an MTA module, the core component of MTA-Net, is introduced on top of a pre-trained CNN architecture to facilitate efficient feature learning focused on lesion regions. The MTA module consists of a sequence of multi-scale triplet spatial attention (MTSA) and multi-scale triplet channel attention (MTCA), which aid in learning detailed feature relationships across spatial and channel dimensions at various scales, enriching feature representations. The proposed MTA-Net is evaluated on two benchmark datasets: HAM10000 and ISIC 2019. The experimental results demonstrate that the MTA-Net outperforms the baseline CNN frameworks and state-of-the-art methods across both datasets while requiring no external data, achieving an accuracy of 91.51% and 78.4%, and a balanced multiclass accuracy (BMCA) of 87.18% and 66.7% on HAM10000 and ISIC 2019 datasets, respectively. Further, the ablation studies and visualization results indicate the potency of the proposed MTA module.