https://atlas.rsna.org/schemas/2025-11/model.json

MRSegmentator

https://doi.org/10.1148/ryai.240777

keno.bressem@tum.de

This work was in large part funded by the Wilhelm Sander Foundation. It was also funded by the European Union (grant no. 101079894). The NAKO is funded by the Federal Ministry of Education and Research (BMBF) (project funding reference no. 01ER1301A/B/C, 01ER1511D, and 01ER1801A/B/C/D), federal states of Germany and the Helmholtz Association, the participating universities and the institutes of the Leibniz Association.

This retrospective study was conducted in accordance with the Declaration of Helsinki and received approval from the local ethics committee (EA4/062/20) with a waiver of patient consent.

This retrospective study developed MRSegmentator, an nnU-Net-based cross-modality deep learning model for multiorgan segmentation of MRI and CT scans, segmenting 40 anatomic structures. It was trained on UK Biobank Dixon MRI, in-house abdominal MRI, and TotalSegmentator CT data, using a human-in-the-loop annotation workflow. The model was validated on NAKO, AMOS22, and TotalSegmentator MRI datasets, demonstrating high class-wise Dice similarity coefficients for well-defined organs and generalizability to CT, outperforming existing open-source tools.

nnU-Net-based cross-modality image segmentation model (nnU-Net version 2)

Code and trained weights are publicly available at https://github.com//hhaentze/MRSegmentator.

Automated segmentation of anatomic structures in medical imaging enables precise organ volumetry, facilitates anatomic context for AI-based diagnosis, and supports quantitative imaging biomarker extraction. It allows researchers to obtain biomarkers relevant for various research questions and clinical tasks.

Diagnosis

Fully automated

Multiorgan segmentation of MRI and CT scans for 40 anatomic structures, including well-defined organs, organs with anatomic variability, and smaller structures. Applicable for whole-body imaging from shoulders to knees.

Three-dimensional (3D) MR or CT images, including UK Biobank Dixon MRI sequences, in-house abdominal MRI sequences, TotalSegmentator CT dataset, NAKO MRI sequences (T1-weighted 3D volumetric interpolated breath-hold examination two-point Dixon sequences (in-phase, opposed-phase, water-only, fat-only, and T2-weighted half-Fourier acquisition single-shot turbo spin echo [HASTE] sequences)), AMOS22 MRI sequences, and TotalSegmentator MRI sequences.

Code and trained weights are publicly available at https://github.com//hhaentze/MRSegmentator. The model was trained with fivefold cross-validation using the 3d_fullres_no_flipping configuration of nnU-Net version 2 with an increased batch size of eight, keeping other training parameters at default values. Preprocessing steps for MRI scans included intensity inversion and histogram equalization.

Human-in-the-loop annotation may have introduced bias, but strong performance on independent datasets suggests minimal impact. UKBB training data represents only 50 unique participants, limiting anatomic variety. Observed sex-based performance differences highlight the need for more balanced training datasets. Model occasionally confused left and right structures, likely due to nnU-Net’s patch-wise design limiting global context. Struggled with postoperative anatomy and some irregular tumor borders.

MRSegmentator is a valuable tool for researchers and clinicians for reproducible, high-accuracy delineation of whole-body anatomy on MRI and CT, offering practical advantages over modality-specific solutions.

Objective cross-modality segmentation with MRSegmentator provided reproducible, high-accuracy delineation of whole-body anatomy on MRI.