A generalisation study in deep learning-based segmentation of lower-limb muscles across different populations

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anatomy

Background:

• Accurate lower-limb muscle segmentation is crucial for clinical applications.
• Anatomical variations across age groups present significant segmentation challenges.
• Existing deep learning models struggle with generalization across diverse populations.

Purpose of the Study:

• To evaluate deep learning models for automatic lower-limb muscle segmentation in typically developed children (TDC).
• To investigate the generalization ability of the Attention-Feature-Fusion-Unet (AFFU) model across different cohorts (healthy young people and post-menopausal women).

Main Methods:

• Manual segmentation and reproducibility analysis of T1-weighted images from TDC.
• Comparison of AFFU model against U-Net, U-Net++, and Attention U-Net.
• Evaluation of model performance using Dice Similarity Coefficient, Relative Volume Error, Hausdorff Distance, and Average Symmetric Surface Distance.

Main Results:

• AFFU model achieved a Dice Similarity Coefficient of 0.86 and Relative Volume Error of 0.09 in the children cohort.
• AFFU significantly reduced Hausdorff Distance (~34%) and Average Symmetric Surface Distance (~20%) compared to U-Net.
• Segmentation accuracy was higher for larger, regular muscles; smaller, irregular muscles were more challenging.
• Single-cohort training limited model generalization; mixed-cohort training improved results.

Conclusions:

• The AFFU model demonstrates superior performance for lower-limb muscle segmentation in children.
• Model generalization requires training on diverse, multi-class cohorts, particularly with complex models employing attention mechanisms.
• Future research should focus on robust segmentation across varied anatomical structures and populations.