https://atlas.rsna.org/schemas/2025-11/model.json

A Deep Learning Algorithm for Automatic 3D Segmentation of Rotator Cuff Muscle and Fat from Clinical MRI Scans

https://dx.doi.org/10.1148/ryai.220132

1.0

Corresponding author: Silvia S. Blemker; email: moc.scitylanakobgnirps@rekmelb.aivlis (obfuscated)

Supported by the National Institutes of Health–National Institute of Arthritis and Musculoskeletal and Skin Diseases (grant 5R41AR078720). Springbok Analytics provided financial support through a federally funded NIH–NIAMS grant.

Both locations were approved by an institutional review board, with a waiver of informed consent, and complied with Health Insurance Portability and Accountability Act guidelines.

Modified 3D U-Net with sliding window; two-stage framework: Stage 1 segments muscle (including fat) and bone boundaries; Stage 2 segments fat infiltration; pixels labeled fat in Stage 2 within Stage 1 muscle are FI ROIs.

Model code is available upon request (authors).

Objective 3D quantification of rotator cuff muscle volumes and intramuscular fat infiltration from clinical MRI; faster than manual segmentation and may aid preoperative assessment and research on RC tears.

Workflow optimization; supports quantitative image analysis for clinical decision making.

Automated 3D segmentation with manual vetting/corrections (approximately 10 minutes per scan vs 4 hours manual).

Quantification of 3D muscle volume and intramuscular fat infiltration of supraspinatus, infraspinatus, teres minor, and subscapularis in adults using clinically acquired sagittal T1-weighted shoulder MRI in both normal rotator cuffs and rotator cuff tears.

Clinically obtained sagittal T1-weighted shoulder MR images; preprocessed with bias correction and resizing.

Preprocess images with bias correction and standardize resolution; run two-stage 3D U-Net segmentation (Stage 1: muscles and bones; Stage 2: fat). Combine outputs to label intramuscular fat within each muscle; review/vet AI output for minor corrections.

Limited and variable sagittal scan coverage inhibits interpretation of absolute muscle volume; small validation dataset; RC tear dataset skewed toward lower Goutallier grades; no external validation; lack of functional outcome data; largest errors in infraspinatus and teres minor due to difficult boundary; Otsu method not ideal for 3D fat quantification.

Use the two-stage AI model for 3D segmentation of RC muscles and intramuscular fat; vet outputs for minor corrections.

Internal validation with interobserver manual segmentations on validation scans; code available upon request.

Processing time reported: AI segmentation plus vetting approximately 10 ± 5 minutes per scan (versus ~4 hours manual).