Segmenting Whole-Body MRI and CT for Multiorgan Anatomic Structure Delineation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate segmentation of anatomical structures in medical imaging is crucial for diagnosis and treatment planning.
• Existing automated segmentation tools often struggle with cross-modality generalization and accuracy for a wide range of organs.

Purpose of the Study:

• To develop and validate MRSegmentator, a novel deep learning model for retrospective, cross-modality multiorgan segmentation of MRI scans.
• To assess the model's performance across diverse MRI sequences and its generalization capabilities to CT scans.

Main Methods:

• Trained MRSegmentator on a large dataset including UK Biobank MRI, in-house abdominal MRI, and TotalSegmentator-CT datasets.
• Utilized a human-in-the-loop annotation workflow with cross-modality transfer learning from a CT segmentation model.
• Evaluated performance on independent MRI datasets (NAKO, AMOS22, TotalSegmentator-MRI) and CT data using Dice Similarity Coefficient (DSC).

Main Results:

• MRSegmentator achieved high Dice Similarity Coefficients (DSC) for well-defined organs (e.g., lungs: 0.81-0.96, heart: 0.81-0.94) and organs with anatomical variability (e.g., liver: 0.82-0.96, kidneys: 0.77-0.95).
• The model demonstrated robust performance on external test datasets, with average DSC ranging from 0.85 to 0.91 for different MRI sequences.
• MRSegmentator generalized effectively to CT scans, achieving a mean DSC of 0.84 on the AMOS CT dataset.

Conclusions:

• MRSegmentator accurately segments 40 anatomical structures in MRI scans.
• The model exhibits strong cross-modality generalization, performing well on both MRI and CT imaging.
• MRSegmentator surpasses existing open-source tools in segmentation accuracy and applicability.