Optimizing MRI annotation workflows for high-accuracy deep learning thigh muscle segmentation in athletes.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Musculoskeletal Imaging

Background:

• Accurate thigh muscle segmentation in MRI is crucial for assessing muscle health.
• Manual segmentation is time-consuming and variable, while automated methods have limitations.
• Defining data requirements for robust automated segmentation is a critical unmet need.

Purpose of the Study:

• Determine the number of annotated lower extremity MRI studies needed for accurate deep learning (DL) thigh muscle segmentation.
• Assess the impact of training dataset size on the agreement of downstream quantitative MRI measures.

Main Methods:

• Generated ground-truth manual segmentations (SegM) for 14 thigh muscles from lower extremity MR images.
• Trained 13 deep learning (nnU-Net) models with increasing training data (Ntrain = 5 to 120).
• Evaluated automated segmentation (SegA) using geometric metrics and compared quantitative MRI measures (fat fraction, diffusion tensor imaging) against manual segmentations.

Main Results:

• Deep learning model training with 20 annotated subjects achieved high accuracy (DSC 0.94 ± 0.02).
• Performance showed modest improvement with 50 training subjects.
• Quantitative measures derived from automated segmentation agreed well with manual segmentations.

Conclusions:

• Twenty annotated MRI images are sufficient for clinically acceptable thigh muscle segmentation performance.
• This finding supports streamlined segmentation and quantitative reporting in athlete care and research.
• Optimized data requirements facilitate the scalability of automated MRI analysis.