HMT-Net: Transformer and MLP Hybrid Encoder for Skin Disease Segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are widely used for skin disease image segmentation.
• CNNs struggle with long-range context, leading to segmentation blur in lesion images.
• A semantic gap exists in CNNs when extracting deep semantic features.

Purpose of the Study:

• To address the limitations of CNNs in skin lesion segmentation.
• To develop a novel hybrid network for improved segmentation accuracy.
• To enhance the network's ability to capture both global and local features.

Main Methods:

• Designed a hybrid encoder network named HMT-Net, integrating Transformer and Multilayer Perceptron (MLP) architectures.
• Utilized the CTrans module's attention mechanism to learn global feature map relevance.
• Employed the TokMLP module with tokenized MLP axial displacement for enhanced boundary and local feature extraction.

Main Results:

• HMT-Net achieved state-of-the-art performance on ISIC2018, ISBI2017, and ISBI2016 datasets.
• Achieved Dice indices of 82.39%, 75.53%, and 83.98%, and IOU of 89.35%, 84.93%, and 91.33%.
• Outperformed the FAC-Net, improving Dice index by up to 1.99% and IOU by up to 2.36%.

Conclusions:

• The proposed HMT-Net effectively addresses the limitations of traditional CNNs in skin lesion segmentation.
• The hybrid architecture enhances the understanding of global lesion information and local boundary details.
• HMT-Net demonstrates superior performance compared to existing state-of-the-art methods.