Automated thigh muscle segmentation using foundation models: validation and impact of segmentation quality

This study aimed to validate a segmentation pipeline for the extraction of volumetric thigh muscle data, investigate the influence of demographics on segmentation performance, and investigate the influence of low-quality segmentations on quantitative performance and downstream metrics.

Manual thigh muscle segmentations on axial T1-weighted spin echo thigh images of 364 (106 training and validation; 258 testing) participants of the Osteoarthritis Initiative (OAI) were used. A trained YOLO (You Only Look Once) object detection model generated labels and box predictions, that were passed to a fine-tuned foundation model (Segment Anything, Meta), which generated muscle segmentations. The quality of segmentations was assessed by two radiologists (none, minor, major issue), and Dice scores and mean bias versus manual segmentations were calculated in the testing dataset. Linear regression was used to calculate the associations of age, sex, and BMI with Dice scores.

Dice scores ranged between 0.97 (quadriceps) and 0.89 (adductors). Exclusion of erroneous segmentations decreased mean bias by 0.06 mL across muscle groups; mean Dice of segmentations with versus without major issues was 0.90 versus 0.94. Female sex was associated with marginally lower Dice scores (beta -0.01, 95% CI: -0.02, -0.01, p < 0.001), while no other demographic feature showed significant associations.

Foundation models can be fine-tuned on a modest task-specific dataset to achieve segmentation accuracy comparable to bespoke networks, without training a dedicated architecture from scratch. Visually erroneous segmentations introduce only negligible degradation of quantitative performance and downstream metrics and may thus be challenging to identify by data-driven detection strategies.

Question Foundation models can be efficiently leveraged for robust segmentation on complex thigh muscle anatomy Findings Radiologist-identified erroneous segmentations introduced only negligible degradation of summary quantitative performance and are thus difficult to identify using statistical methods. Relevance statement Human-in-the-loop quality assurance post-validation adds limited value to improve downstream volumetric muscle metrics in foundation-model-based thigh segmentation. These findings carry implications for both research and potential clinical applications of muscle volume quantification, e.g., personalized therapeutic regimens.