CyCoSeg: A Cyclic Collaborative Framework for Automated Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks excel at general image segmentation but struggle with medical images due to small object sizes.
• Accurate medical image segmentation is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To develop a robust framework, CyCoSeg, that overcomes limitations in medical image segmentation, particularly for small or challenging objects.
• To improve the performance of existing semantic segmentation networks (SSNs) without major architectural changes.

Main Methods:

• Proposed a cyclic collaborative framework, CyCoSeg, integrating a deep active shape model (D-ASM) with a semantic segmentation network (SSN).
• Implemented an Expectation Maximization formulation for SSN to guide D-ASM keypoint identification.
• Utilized D-ASM segmentation proposals to guide the SSN, iterating until convergence.

Main Results:

• CyCoSeg significantly boosts the performance of baseline segmentation models, including popular SSNs.
• Achieved highly competitive results in left ventricle segmentation accuracy on benchmark datasets.
• Demonstrated effective generalization for segmenting lungs and kidneys in CT scans.

Conclusions:

• CyCoSeg offers an effective solution for challenging medical image segmentation tasks, especially those involving small objects.
• The collaborative cyclic framework enhances segmentation accuracy and model performance without requiring substantial architectural modifications.
• The method shows strong potential for various clinical applications requiring precise organ segmentation.