Deep-Learning-Based Segmentation of the Shoulder from MRI with Inference Accuracy Prediction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedic Surgery

Background:

• Three-dimensional (3D) imaging aids rotator cuff tear prognosis.
• Clinical application requires efficient and robust MRI segmentation.
• Manual segmentation is time-consuming and limits 3D diagnostic use.

Purpose of the Study:

• To develop an automated deep learning method for segmenting shoulder anatomy from MRI.
• To integrate automatic verification of segmentation accuracy.
• To facilitate clinical adoption of 3D image-based diagnosis for rotator cuff tears.

Main Methods:

• Utilized a deep learning network (nnU-Net) for automatic segmentation of humerus, scapula, and rotator cuff muscles.
• Trained the network on 111 T1-weighted MRI scans and tested on 60 scans from 76 patients.
• Adapted the nnU-Net framework to estimate label-specific network uncertainty for automatic result verification.

Main Results:

• Achieved an average Dice coefficient of 0.91 ± 0.06 for anatomical segmentation.
• The uncertainty estimation successfully identified segmentations requiring correction with 1.0 sensitivity and 0.94 specificity.
• Automated verification significantly reduced the need for manual review.

Conclusions:

• The presented deep learning approach enables efficient and accurate automatic segmentation of shoulder anatomy from MRI.
• Integrated uncertainty estimation provides robust automatic verification of segmentation quality.
• These methods streamline 3D image analysis, making it practical for routine clinical use in rotator cuff tear management.