Dual-branch Transformer for semi-supervised medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Medical image segmentation is crucial but hindered by time-consuming annotations and limited labeled data.
• Current convolutional neural networks often fail to capture both local and global image features, limiting segmentation of complex structures.

Purpose of the Study:

• To propose a novel deep learning network architecture for medical image segmentation.
• To overcome challenges associated with limited labeled data and insufficient feature representation in existing models.

Main Methods:

• A U-shaped encoder-decoder structure with dual-branch encoders utilizing Swin modules (shift window) for multi-scale feature capture.
• A semi-supervised learning strategy incorporating unlabeled data via a level set function for consistency between regression and classification.

Main Results:

• Achieved 74.56% segmentation accuracy on the COVID-19 CT dataset.
• Attained 79.79% segmentation accuracy on the DRIVE dataset.
• Demonstrated superior performance compared to various semi-supervised and fully supervised models.

Conclusions:

• The proposed method effectively enhances feature extraction using Swin modules with varied window sizes.
• The level set function significantly improves the utilization of unlabeled data in semi-supervised medical image segmentation.
• The findings offer valuable insights for advancing deep learning applications in medical imaging.