Reducing annotation burden in MR: A novel MR-contrast guided contrastive learning approach for image segmentation

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Deep learning (DL) models benefit from contrastive learning for pretraining, especially in medical image segmentation with limited annotations.
• Learning domain-specific local representations is crucial for improving DL model performance in such scenarios.

Purpose of the Study:

• To extend contrastive learning for Magnetic Resonance (MR) image segmentation using unlabeled data.
• To leverage domain-specific contrast information to enhance downstream MR image segmentation tasks with limited labeled data.

Main Methods:

• Proposed a novel constrained contrastive learning (CCL) strategy using tissue-specific information via constraint maps.
• Defined positive and negative local neighborhoods for contrastive learning, embedding tissue properties into the representational space.
• Demonstrated utility in multi-organ segmentation (T2-weighted images) and tumor segmentation (multi-parametric BraTS dataset).

Main Results:

• CCL strategy consistently improved Dice scores, Precision, and Recall across all segmentation tasks.
• Observed performance comparable to supervised baselines with reduced annotation effort.
• t-SNE visualization confirmed embedding of T2 information; multi-contrast pretraining further improved BraTS segmentation.

Conclusions:

• Embedding tissue-specific information via CCL enhances DL model performance in MR image segmentation.
• CCL offers a promising approach to improve performance and address data scarcity in medical image segmentation.