MedScale-Former: Self-guided multiscale transformer for medical image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate medical image segmentation is vital for automated clinical decisions.
• Supervised deep learning methods require extensive labeled data, posing a significant challenge.
• Existing methods struggle with data scarcity and boundary definition.

Purpose of the Study:

• To develop a novel self-supervised deep learning approach for medical image segmentation.
• To overcome the limitations of data-hungry supervised methods.
• To improve segmentation accuracy and boundary definition without extensive manual labeling.

Main Methods:

• A dual-branch transformer network operating on two scales to capture global and local information.
• Leveraging semantic dependencies between scales for inter-scale consistency (self-supervised learning).
• Incorporating spatial stability loss for self-supervised content clustering and a selective kernel regional attention module for boundary refinement.

Main Results:

• Demonstrated superior performance in segmentation tasks compared to state-of-the-art methods.
• Achieved accurate segmentation of skin lesions, lung organs, and multiple myeloma plasma cells.
• The proposed method effectively reduces the need for large labeled datasets.

Conclusions:

• The novel dual-branch transformer network offers a powerful self-supervised approach for medical image segmentation.
• This method significantly enhances segmentation accuracy and boundary definition.
• It provides a viable solution for clinical applications where labeled data is scarce.