Two-Stage Self-Supervised Contrastive Learning Aided Transformer for Real-Time Medical Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• High-quality annotated medical datasets are scarce, hindering segmentation task performance.
• Self-supervised pre-training, especially contrastive learning, offers a solution by leveraging unlabeled data.
• Existing methods struggle with limited annotated volumes in medical image segmentation.

Purpose of the Study:

• To develop a self-supervised semantic segmentation framework for medical images with limited annotations.
• To enhance the distinction of anatomical regions using intrinsic data similarities.
• To improve segmentation performance through a novel parallel transformer module.

Main Methods:

• A self-supervised fusion network combining segmentation and contrastive loss for anatomical region distinction.
• An efficient parallel transformer module utilizing multi-view multiscale feature fusion and depth-wise features.
• A multi-encoder transformer architecture trained self-supervisedly, then fine-tuned with limited annotated data for brain tumor segmentation.

Main Results:

• The proposed multi-encoder transformer model significantly improved outcomes across three medical image segmentation tasks.
• The framework effectively leverages intrinsic anatomical similarities for enhanced segmentation.
• The solution outperformed state-of-the-art methodologies when validated on diverse medical image datasets.

Conclusions:

• Self-supervised learning combined with a fusion network and parallel transformer module offers a powerful solution for medical image segmentation with limited annotations.
• The proposed method effectively enhances anatomical feature distinction and achieves superior performance.
• This approach holds significant potential for advancing medical image analysis and clinical applications.