Multicenter, Multivendor Development and Evaluation of Automated Liver MR Image Prescription

Motivation: Manual image prescription is susceptible to human error, limiting efficiency and degrading prescription reproducibility. Previous work addressing this relies on single-center, single-vendor datasets. Goal(s): Extend previous work on automated liver MR image prescription, now with performance validation across vendors, center, field strength, and known liver pathology. Approach: Train and evaluate a YOLOv8 model on a large manually curated and annotated dataset (11012 total; 7707 training, 1102 validation, 2203 test) spanning a wide variety of acquisition and patient characteristics that may impact performance. Results: Our model was successful in prescribing liver MR volumes with high performance across vendors, field strength, and known liver pathology. Impact: The validated multicenter, multivendor liver prescription model may improve liver MRI workflow efficiency and reproducibility. Further, the trained models will be made available upon publication, providing an important resource for research and clinical MR applications in the liver.