UTADC-Net: Unsupervised Topological-Aware Diffusion Condensation Network for Medical Image Segmentation

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Computational anatomy

Background:

• Medical image segmentation is vital for diagnosis and treatment planning.
• Unsupervised methods offer promise for clinical applications by using unlabeled data.
• Existing unsupervised methods struggle with maintaining anatomical topological consistency, leading to errors.

Purpose of the Study:

• To introduce a novel unsupervised network for medical image segmentation.
• To enhance topological consistency and structural integrity in segmentation results.
• To provide a practical solution for unsupervised medical image segmentation.

Main Methods:

• Developed the Unsupervised Topological-Aware Diffusion Condensation Network (UTADC-Net).
• Employed a pixel-centric patch embedding module for fusing local and global information.
• Introduced an adaptive topological constraint mechanism for learning anatomically aligned representations.

Main Results:

• UTADC-Net significantly outperformed existing unsupervised methods on three datasets.
• Achieved superior segmentation accuracy and topological structure preservation.
• Demonstrated excellent anatomical structural consistency in segmentation results.

Conclusions:

• The proposed UTADC-Net offers a novel and practical solution for unsupervised medical image segmentation.
• The diffusion condensation framework effectively models long-range dependencies and incorporates topological constraints.
• The method successfully addresses limitations of existing unsupervised segmentation techniques.