Positional contrastive learning for improved thigh muscle segmentation in MR images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate segmentation of thigh muscles is crucial for quantitative MRI analysis.
• Deep learning models require extensive labeled data, which is time-consuming to obtain for muscle segmentation.
• Self-supervised learning (SSL) offers a solution by leveraging unlabeled data for model pretraining.

Purpose of the Study:

• To propose and evaluate positional contrastive SSL for segmenting individual thigh muscles in MRI scans.
• To assess the effectiveness of SSL pretraining with varying amounts of limited annotated data.
• To compare SSL performance against a randomly initialized model for thigh muscle segmentation.

Main Methods:

• Utilized a U-Net architecture pretrained with positional contrastive SSL on unlabeled T1w MRI thigh acquisitions.
• Compared SSL pretraining with a randomly initialized model (RND) on a labeled dataset.
• Evaluated segmentation performance using Dice Similarity Coefficient (DSC) with increasing numbers of annotated volumes (1 to 40).

Main Results:

• SSL pretraining significantly improved thigh muscle segmentation accuracy, particularly with very limited labeled data (e.g., DSC 0.631 vs. 0.530 for 1 labeled volume).
• Substantial enhancements were observed even when using the full labeled dataset (DSC 0.927 for SSL vs. 0.924 for RND).
• Positional contrastive SSL demonstrated effectiveness in improving segmentation with minimal annotated data.

Conclusions:

• Positional contrastive SSL is an effective method for accurate thigh muscle segmentation from MRI.
• SSL significantly reduces the need for extensive manual annotation, potentially accelerating clinical workflows.
• The approach shows promise for quantitative MRI analysis in populations like elderly healthy subjects.