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

Area of Science:

• Dermatology
• Computer Vision
• Artificial Intelligence

Background:

• Multiclass classification of skin lesions is vital for computer-aided skin cancer diagnosis.
• High similarity between lesion classes and intra-class variations pose significant challenges.
• Existing Convolutional Neural Network (CNN) models struggle to capture subtle, multi-scale lesion details.

Purpose of the Study:

• To propose a novel Multi-scale Triplet Attention-aware Network (MTA-Net) for enhanced multiclass skin lesion classification.
• To improve the learning of fine-grained lesion information at various scales.
• To address the limitations of current CNN and attention-driven models in capturing subtle lesion variations.

Main Methods:

• Developed MTA-Net, incorporating a Multi-scale Triplet Attention (MTA) module atop a pre-trained CNN.
• The MTA module utilizes Multi-scale Triplet Spatial Attention (MTSA) and Multi-scale Triplet Channel Attention (MTCA).
• Evaluated MTA-Net on the HAM10000 and ISIC 2019 benchmark datasets.

Main Results:

• MTA-Net achieved superior performance compared to baseline CNNs and state-of-the-art methods on both datasets.
• Achieved 91.51% accuracy and 87.18% Balanced Multiclass Accuracy (BMCA) on HAM10000.
• Achieved 78.4% accuracy and 66.7% BMCA on ISIC 2019, without external data.

Conclusions:

• MTA-Net effectively captures detailed feature relationships across spatial and channel dimensions at various scales.
• The proposed MTA module significantly enhances feature representations for improved classification.
• MTA-Net demonstrates strong potential for advancing computer-aided skin cancer diagnosis.