Swin-Net: A Swin-Transformer-Based Network Combing with Multi-Scale Features for Segmentation of Breast Tumor...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Breast cancer is a leading global cancer, particularly among women.
• Accurate breast tumor segmentation is crucial for diagnosis and treatment planning.
• Existing segmentation methods struggle with the complexities of ultrasound image acquisition and tumor characteristics.

Purpose of the Study:

• To introduce Swin-Net, a novel semantic segmentation framework for breast ultrasound images.
• To improve the accuracy and robustness of breast tumor segmentation.
• To leverage the global modeling capabilities of Swin-Transformer combined with CNNs.

Main Methods:

• Developed Swin-Net, integrating a Swin-Transformer encoder with CNNs.
• Introduced a feature refinement and enhancement module (RLM) to refine learned features.
• Implemented a hierarchical multi-scale feature fusion module (HFM) for cross-layer fusion and noise suppression.

Main Results:

• Swin-Net demonstrated superior performance compared to state-of-the-art methods on public datasets.
• Achieved an absolute improvement of 1.4-1.8% in Dice coefficient.
• Validated the model's effectiveness on a newly introduced breast ultrasound image dataset.

Conclusions:

• Swin-Net significantly advances breast ultrasound image segmentation.
• The proposed framework offers valuable exploration for research and clinical applications in breast cancer detection.
• The integration of Transformer and CNNs with specialized modules enhances segmentation accuracy.