MAS-CL: An End-to-End Multi-Atlas Supervised Contrastive Learning Framework for Brain ROI Segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Accurate brain region-of-interest (ROI) segmentation is crucial for neurological analysis.
• Deep learning for brain ROI segmentation is hindered by insufficient annotated data.

Purpose of the Study:

• To propose a novel multi-atlas supervised contrastive learning framework (MAS-CL) for end-to-end brain ROI segmentation using MR images.
• To address the challenge of limited annotated data in deep learning-based brain segmentation.

Main Methods:

• Developed a two-step MAS-CL framework: 1) multi-atlas supervised contrastive learning for latent representation using limited voxel-level labeled MR images, and 2) brain ROI segmentation using the pre-trained backbone.
• Utilized multi-atlas supervised information to pre-train the backbone, defining sample pair correlations via label maps of input and atlas images.

Main Results:

• The MAS-CL framework was evaluated on five diverse datasets (LONI-LPBA40, IXI, OASIS, ADNI, CC359).
• Experimental results demonstrated significant improvements in brain ROI segmentation performance across all tested datasets.
• The proposed method effectively leverages limited labeled data for enhanced segmentation.

Conclusions:

• The MAS-CL framework offers a robust solution for brain ROI segmentation from MR images, particularly in low-data regimes.
• This approach significantly enhances segmentation accuracy, paving the way for more reliable brain analysis.
• The study highlights the potential of multi-atlas supervised contrastive learning in medical image segmentation.