TC-Net: Dual coding network of Transformer and CNN for skin lesion segmentation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Skin lesion segmentation is crucial for medical AI applications.
• Convolutional Neural Networks (CNNs) excel at local feature extraction but struggle with global context.
• Transformers are effective for global context but lack fine-grained local feature extraction capabilities.

Purpose of the Study:

• To develop an improved deep learning architecture for precise skin lesion segmentation.
• To leverage the complementary strengths of CNNs and Transformers for enhanced performance.
• To address limitations of existing models in handling complex, low-contrast skin images.

Main Methods:

• Proposed a novel dual coding fusion network architecture named TC-Net.
• Integrated CNNs for local feature extraction and Transformers for global feature modeling.
• Evaluated TC-Net on multiple skin image datasets, including ISIC2018 and ISBI2017.

Main Results:

• TC-Net demonstrated significant improvements in global segmentation performance compared to single-network models.
• Achieved a 2.46% increase in the Dice index and approximately 4% in the JA index on ISIC2018 compared to Swin UNet.
• Showcased approximately 4% improvement in both Dice and JA indices on the ISBI2017 dataset.

Conclusions:

• The fusion of CNNs and Transformers in TC-Net effectively combines local and global feature information for superior skin image segmentation.
• TC-Net exhibits robustness and outperforms existing single-network models, offering a promising approach for medical image analysis.
• The findings highlight the potential of hybrid architectures for advancing the accuracy and reliability of automated skin lesion detection.