Semi-automatic muscle segmentation in MR images using deep registration-based label propagation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automated muscle segmentation using convolutional neural networks (CNNs) shows promise in magnetic resonance (MR) imaging but requires extensive training data.
• Manual segmentation of muscles in pediatric and rare disease cohorts is time-consuming and labor-intensive, especially for 3D volumes.

Purpose of the Study:

• To develop a novel 3D muscle segmentation method that utilizes a limited number of annotated 2D slices.
• To improve the efficiency and accuracy of muscle segmentation for rare disease and pediatric populations.

Main Methods:

• Proposed a registration-based label propagation technique for 3D muscle segmentation.
• Employed an unsupervised deep registration scheme to ensure anatomical consistency.
• Implemented a loss function that penalizes inconsistent segmentation across annotated slices.

Main Results:

• The developed few-shot multi-label segmentation model achieved superior performance compared to state-of-the-art methods.
• Demonstrated effectiveness on MR data from lower leg and shoulder joints.
• Significantly reduced the need for extensive manual annotations.

Conclusions:

• The proposed method offers an efficient solution for 3D muscle segmentation with limited annotated data.
• This approach has the potential to accelerate research in pediatric and rare disease muscle imaging.
• Registration-based label propagation is a viable strategy for few-shot medical image segmentation.