Benchmarking Multi-Organ Segmentation Tools for Multi-Parametric T1-weighted Abdominal MRI.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Multi-organ segmentation in multi-parametric MRI is crucial for correlating imaging biomarkers with disease status.
• Publicly available tools like MRSegmentator (MRSeg), TotalSegmentator MRI (TS), and TotalVibeSegmentator (VIBE) exist for MRI segmentation.
• Performance variations of these tools across different MRI sequences are not well-quantified.

Purpose of the Study:

• To benchmark the performance of three public multi-organ segmentation tools (MRSeg, TS, VIBE) on specific MRI sequence types.
• To evaluate the accuracy of these tools for abdominal structure segmentation in a curated dataset.

Main Methods:

• A subset of 40 multi-parametric MRI volumes from the Duke Liver Dataset was curated, including 10 volumes each of pre-contrast fat saturated T1, arterial T1w, venous T1w, and delayed T1w phases.
• Ten abdominal structures were manually annotated.
• The performance of MRSeg, TS, and VIBE was evaluated using Dice score and Hausdorff Distance (HD) error.

Main Results:

• MRSegmentator (MRSeg) achieved a Dice score of 80.7 ± 18.6 and a Hausdorff Distance (HD) error of 8.9 ± 10.4 mm.
• MRSeg demonstrated significantly better performance (p < .05) across the evaluated MRI sequence types compared to TS and VIBE.

Conclusions:

• MRSegmentator (MRSeg) is a superior tool for multi-organ segmentation in the evaluated multi-parametric MRI sequences.
• This benchmarking provides critical insights for selecting appropriate segmentation tools in radiological applications.