Multicenter, Multivendor Development and Validation of Automated Liver Prescription.
Authors
Affiliations (11)
Affiliations (11)
- Department of Radiology, University of Wisconsin-Madison, Madison, Wisconsin, USA.
- Department of Medical Physics, University of Wisconsin-Madison, Madison, Wisconsin, USA.
- GE HealthCare, Waukesha, Wisconsin, USA.
- Department of Radiology, Duke University Medical Center, Durham, North Carolina, USA.
- Department of Radiology, University of Texas Southwestern Medical Center, Dallas, Texas, USA.
- Department of Medicine, Duke University Medical Center, Durham, North Carolina, USA.
- Center for Advanced Magnetic Resonance Development, Duke University Medical Center, Durham, North Carolina, USA.
- Department of Biomedical Engineering, University of Wisconsin-Madison, Madison, Wisconsin, USA.
- Department of Medicine, University of Wisconsin-Madison, Madison, Wisconsin, USA.
- Department of Emergency Medicine, University of Wisconsin-Madison, Madison, Wisconsin, USA.
- Department of Electrical and Computer Engineering, University of Wisconsin-Madison, Madison, Wisconsin, USA.
Abstract
Manual prescription of liver MRI volumes introduces inter-operator variability and prolongs exam duration. Artificial intelligence (AI)-based prescription has been demonstrated, but existing work has been limited to single-vendor settings. To develop and validate an AI model for automated liver MRI prescription using a large, consecutive, multicenter dataset encompassing multiple vendors, field strengths, localizer sequences, and liver pathologies. Retrospective. A total of 11,012 patients (5928 female) who underwent liver MRI exams across three vendors (GE HealthCare, Philips Healthcare, Siemens Healthineers) at three institutions, split into training/validation/test sets (70/10/20%). Liver disorders included cirrhosis, ascites, portal hypertension, and iron overload. Three-plane localizers acquired using single-shot fast spin echo (68.3%) and gradient echo (31.7%) at 1.5 T (52.8%) and 3 T (47.2%). A YOLOv8 object detection model was trained to detect the liver, torso, and arms on localizer images. Automated prescription accuracy was assessed using three-dimensional intersection-over-union (IoU<sub>3D</sub>) for axial, coronal, and sagittal prescriptions. Subgroup analyses evaluated performance across technical and patient-related factors. Mann-Whitney U tests with Bonferroni correction, Kruskal-Wallis tests, Wilcoxon signed-rank tests, Spearman correlation. Effect sizes were calculated using rank-biserial correlation (r<sub>rb</sub>) and epsilon-squared (ε<sup>2</sup>). p < 0.05 was considered statistically significant. Median axial prescription IoU<sub>3D</sub> was 0.963 (interquartile range: 0.946-0.976), with median IoU<sub>3D</sub> above 0.940 for coronal and sagittal prescriptions. For axial prescriptions, margins less than 4 mm in each direction were required to achieve complete coverage of manual prescriptions in 90% of cases in the held-out test set. Effects were negligible to small across vendors, field strengths, localizer sequences, and patient factors (r<sub>rb</sub> ≤ 0.16, ε<sup>2</sup> ≤ 0.01). A single multivendor model achieved accurate automated liver MRI prescription without clinically meaningful performance differences across vendors, field strengths, localizer sequences, or patient factors. 2. Stage 1.