Local contrastive loss with pseudo-label based self-training for semi-supervised medical image segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Supervised deep learning excels in medical image segmentation but demands extensive expert-labeled datasets, which are costly and time-consuming to acquire.
• Semi/self-supervised learning methods mitigate this by utilizing unlabeled data alongside limited annotations.
• Existing self-supervised methods often focus on global representations, with local contrastive learning showing limited success due to reliance on non-semantic features.

Purpose of the Study:

• To develop an effective local contrastive loss for learning robust pixel-level features crucial for medical image segmentation.
• To improve segmentation accuracy in scenarios with limited labeled medical imaging data.
• To leverage semantic information from pseudo-labels of unlabeled data within a self-training framework.

Main Methods:

• Proposed a novel local contrastive loss function that utilizes pseudo-labels from unlabeled images and ground truth labels from limited annotated images.
• The contrastive loss encourages similar representations for pixels with the same semantic label (pseudo or ground truth) and dissimilar representations for pixels with different labels.
• Employed a pseudo-label-based self-training strategy, jointly optimizing the proposed contrastive loss on both labeled and unlabeled data, and a segmentation loss on labeled data.

Main Results:

• Achieved high segmentation performance on three public medical datasets (cardiac and prostate anatomies) using only one or two labeled 3D volumes.
• Demonstrated substantial improvements compared to state-of-the-art semi-supervised, data augmentation, and concurrent contrastive learning methods.
• The proposed method effectively learns discriminative pixel-level features by incorporating semantic information.

Conclusions:

• The proposed local contrastive loss, combined with pseudo-labeling and self-training, significantly enhances medical image segmentation accuracy with minimal labeled data.
• This approach offers a viable solution for overcoming the data annotation bottleneck in medical image analysis.
• The method's effectiveness is validated across diverse anatomical datasets, highlighting its generalizability.