Self-supervised contrastive learning with random walks for medical image segmentation with limited annotations.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Supervised deep learning for medical image segmentation relies heavily on large annotated datasets.
• Self-supervised pre-training strategies aim to reduce the dependency on annotated data.
• Contrastive learning on dense pixel-wise representations is an effective self-supervised tool.

Purpose of the Study:

• To develop a semi-supervised semantic segmentation approach for medical imaging that leverages inherent anatomical similarities.
• To reduce the requirement for annotated training data in medical image segmentation tasks.
• To improve the efficiency and accuracy of segmenting anatomical structures with limited annotations.

Main Methods:

• Adapted cyclical contrastive random walks (CCRW) for self-supervision on paired embedded image slices.
• Employed a contrastive loss alongside a segmentation loss in a single training stage.
• Utilized multi-level supervision to capture both local and global anatomical characteristics.
• Bypassed the need for negative samples by using paths of cyclical random walks.

Main Results:

• Achieved a median increase of 8.01 and 5.90 percentage points in Dice Similarity Coefficient (DSC) compared to baseline across three MRI datasets.
• Demonstrated effectiveness in reducing the amount of required annotations for semantic segmentation.
• Successfully differentiated salient anatomical regions using contrastive learning with limited annotations.

Conclusions:

• The proposed semi-supervised approach effectively reduces annotation requirements for medical image segmentation.
• Leveraging anatomical similarities through adapted CCRW and multi-level supervision enhances segmentation performance.
• This method shows significant promise for improving the practical application of deep learning in medical imaging analysis.