Open-source AI for MRI Segmentation Wins 2025 Margulis Award - Here's Why It Matters
An open-source, sequence-agnostic MRI segmentation model has earned the 2025 Alexander R. Margulis Award for the best original scientific article in Radiology. The work addresses a long-standing bottleneck in radiology: manual segmentation that consumes hours of expert time.
The standout idea is simple and practical-build a model that doesn't depend on a specific MRI sequence. That means one tool that works across protocol variations, without retraining for each setting.
What makes this study different
The researchers developed TotalSegmentator MRI on top of nnU-Net, known for adapting well to new datasets with minimal manual tuning. The model generalizes across MRI scans with different contrast, section thickness, field strength, and pulse sequences.
It brings the ease many clinicians already experienced with TotalSegmentator for CT to MRI-without locking users into sequence-specific training pipelines.
How the model was built
- Training data: 616 MRI and 527 CT scans.
- Anatomy: 80 structures identified for segmentation.
- Resolutions: 1.5-mm isotropic for the main model; a second model (TotalSegmentator MRI-3) at 3-mm isotropic.
- Architecture: nnU-Net-based, with the 80 targets split into six submodels to manage memory efficiently.
- Diverse MRI inputs: varying contrast, slice thickness, field strengths, and sequences.
Performance at a glance
TotalSegmentator MRI closely matched expert-drawn segmentations across all 80 structures on the internal test set. Accuracy was even higher across the 50 most clinically relevant structures.
On comparable evaluations, it outperformed MRSegmentator and AMOS. Thanks to the diverse training data, it kept performance steady across a wide range of MRI types.
A useful surprise: training on CT boosted MRI results
Mixing CT with MRI during training improved MRI segmentation quality. In effect, CT acted like data augmentation, helping the model generalize better to MRI variability.
Clinical impact you can use now
- Reduces manual workload and inter-reader variability.
- Speeds up organ volumetry, treatment planning, and opportunistic screening.
- Enables population-scale studies that would be impractical to do manually.
To show scale, the team processed 8,672 abdominal MRI scans to map age-related organ volume changes. Findings matched clinical expectations: kidney, liver, and spleen volumes declined with age, while adrenal gland volumes increased.
Open science and what's next
The group is releasing the full model, training data, and annotations. Community adoption is already strong, with frequent new tools benchmarking against this work.
Next steps focus on finer structures-peripheral vessels and small muscle groups-to extend clinical utility further.
Technical snapshot
- Framework: nnU-Net
- Training: 616 MRI + 527 CT
- Targets: 80 anatomic structures (split into six submodels)
- Resolutions: 1.5-mm iso; 3-mm iso variant (MRI-3)
- Comparators: MRSegmentator, AMOS
Resources
Award note: The Margulis Award will be presented during RSNA 2025 in Chicago, Nov. 30-Dec. 4.
Building skills for medical imaging AI or clinical data science? Explore curated AI course paths by job role at Complete AI Training.
Your membership also unlocks:
