Automatic field-of-view planning for magnetic resonance shoulder imaging using Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate prescription of oblique coronal and sagittal fields of view (FOV) is crucial for diagnostic shoulder MRI.
• Manual planning is operator-dependent, leading to variability and inconsistent image quality.
• Deep learning (DL) has potential to automate and standardize FOV prescription for shoulder MRI.

Purpose of the Study:

• To develop and evaluate a DL-based automated approach for oblique FOV prescription in shoulder MRI.
• To compare the performance of the DL model against radiographer manual planning.
• To assess the generalizability and clinical utility of the automated system across multiple institutions.

Main Methods:

• A retrospective multicenter study included 575 shoulder MRI examinations.
• A two-stage DL pipeline using YOLOv11-OBB variants was implemented for slice selection and FOV prescription.
• Performance was evaluated using mean absolute slice difference (MASD), intersection over union (IoU), and mean absolute angle difference (MAAD), compared to radiographer prescriptions.

Main Results:

• The YOLOv11-OBB-l model achieved the lowest MASD for slice selection (1.016 slices).
• The YOLOv11-OBB-x model demonstrated superior performance in FOV prescription (coronal IoU: 0.847, sagittal IoU: 0.852, MAAD: 3.259°).
• The DL approach showed non-inferior performance to interrater variability across all sites and metrics, with a mean clinical utility of 97.2%.

Conclusions:

• DL-based automated FOV prescription for shoulder MRI is effective and comparable to radiographers.
• The automated system generalizes well across different institutions.
• This approach offers high clinical utility, potentially improving standardization and workflow in shoulder MRI.