RCPS: Rectified Contrastive Pseudo Supervision for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Fully-supervised medical image segmentation demands extensive expert annotations, which are costly and time-consuming.
• Semi-supervised methods leverage unlabeled data but struggle with noisy pseudo-labels and poor feature separability.
• Existing semi-supervised approaches often exhibit suboptimal performance due to these challenges.

Purpose of the Study:

• To develop an effective semi-supervised medical image segmentation method that overcomes limitations of current approaches.
• To improve the robustness and accuracy of segmentation models using limited labeled data.
• To enhance feature representation for better class discrimination in medical images.

Main Methods:

• Proposed Rectified Contrastive Pseudo Supervision (RCPS) combining rectified pseudo-supervision and voxel-level contrastive learning.
• Implemented a novel rectification strategy using uncertainty estimation and consistency regularization to denoise pseudo-labels.
• Introduced a bidirectional voxel contrastive loss to promote intra-class consistency and inter-class separability.

Main Results:

• RCPS demonstrated superior segmentation performance compared to state-of-the-art methods on public and clinical datasets.
• The rectification strategy effectively reduced noise influence in pseudo-labels.
• Voxel-level contrastive learning significantly improved class separability in the feature space.

Conclusions:

• RCPS offers a robust and effective solution for semi-supervised medical image segmentation.
• The proposed method addresses key challenges in learning from unlabeled medical images.
• RCPS achieves improved segmentation accuracy, making it a valuable tool for clinical applications.