Boundary-aware information maximization for self-supervised medical image segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Self-supervised representation learning enhances model performance on tasks with limited labeled data.
• Contrastive learning is a popular self-supervised method but faces challenges in image segmentation due to difficulties in defining positive and negative pairs for dense feature maps.
• Existing methods like Information Invariant Clustering (IIC) for unsupervised learning in segmentation have limitations in optimization, cluster consistency, and boundary representation.

Purpose of the Study:

• To propose a novel self-supervised pre-training method for image segmentation that overcomes the limitations of contrastive learning and IIC.
• To enhance the learning of local image representations for segmentation tasks.
• To improve the effectiveness of self-supervised learning in medical image segmentation.

Main Methods:

• Leveraging Information Invariant Clustering (IIC) for unsupervised learning of local image representations.
• Introducing a regularized mutual information maximization objective for balanced and consistent clusters across image transformations.
• Proposing a boundary-aware loss function based on cross-correlation to improve region representation.

Main Results:

• The proposed method effectively learns local representations for image segmentation without requiring positive and negative pairs.
• Experimental results on four medical image segmentation tasks demonstrate significant performance improvements over state-of-the-art self-supervised and semi-supervised methods.
• The method achieves performance close to full supervision with minimal labeled data.

Conclusions:

• The novel self-supervised approach significantly advances the field of image segmentation, particularly in data-scarce scenarios.
• The method offers enhanced interpretability through cluster visualization.
• This work provides a powerful tool for medical image segmentation, reducing the need for extensive manual annotation.