Semi-Supervised Medical Image Segmentation Guided by Bi-Directional Constrained Dual-Task Consistency.

Area of Science:

• Medical Image Analysis
• Computer-Aided Surgery
• Machine Learning for Healthcare

Background:

• Multi-object segmentation in medical imaging is crucial for surgical planning, robot-assisted surgery, and safety.
• Pelvic CT segmentation is challenging due to low tissue contrast and limited annotated data.

Purpose of the Study:

• To develop an automatic segmentation algorithm for pelvic CT.
• To leverage unlabeled data to improve segmentation quality in low-contrast medical images.

Main Methods:

• Proposed a bi-direction constrained dual-task consistency model named PICT.
• Employed image interpolation consistency at pixel, model, and data levels to learn from unlabeled data.
• Introduced an auxiliary pseudo-supervision task to constrain interpolation error and ensure reliable predictions.

Main Results:

• PICT achieved high mean DSC scores: 87.18% (ACDC), 96.42% (CTPelvic1k), and 79.41% (Multi-tissue Pelvis).
• Demonstrated performance gains of 0.8%, 0.5%, and 1% over state-of-the-art semi-supervised methods.
• Showed 3-9% improvement compared to baseline supervised methods.

Conclusions:

• The PICT model effectively utilizes unlabeled data to enhance segmentation of low-contrast medical images.
• Improved segmentation accuracy can lead to more precise surgical path planning.
• PICT provides valuable input for robot-assisted surgical systems.