TotalSegmentator MRI: Broad MRI Segmentation

• TotalSegmentator MRI is a sequence-independent segmentation model that extends CT-based TotalSegmentator to 59 anatomical structures on diverse MR images.
• It employs a robust training strategy combining heterogeneous MRI data from multiple scanners and protocols with CT inputs to enhance anatomical supervision.
• Empirical evaluations show strong performance with Dice scores up to 0.824 on MR tests while retaining high accuracy on CT, demonstrating its practical versatility.

TotalSegmentator MRI denotes the MRI extension of the earlier CT-focused TotalSegmentator framework: an open-source, nnU-Net–based system intended for automatic segmentation on routine clinical MR images across heterogeneous MRI sequences rather than a single protocol (D'Antonoli et al., 2024). In the paper’s framing, its role is to extend broad multi-structure segmentation from CT into MRI for use cases such as organ volumetry, disease characterization, surgical planning, treatment planning or disease monitoring, and opportunistic screening. A notable bibliographic ambiguity is that the abstract states segmentation of “80 anatomic structures,” whereas the detailed model description supports 59 anatomical structures; the latter is the consistent label count in the technical summary and category breakdown (D'Antonoli et al., 2024).

1. Definition, scope, and anatomical coverage

TotalSegmentator MRI is presented as a sequence-independent MRI segmentation model. In this context, sequence independence is not implemented through an explicit MRI physics model or a sequence-conditioned architecture. Rather, it is achieved empirically through broad training diversity: MRI series were randomly sampled from routine PACS studies, many different MRI sequences and protocols were included, and CT was added to reinforce anatomy learning beyond MRI appearance variability (D'Antonoli et al., 2024).

The model is broader than narrow abdominal or protocol-specific MRI tools. Its label space spans organs, vessels, bones, muscles, and tissue compartments, and includes structures such as brain, spinal cord, heart, bladder, prostate, vertebrae, intervertebral discs, paired long bones, paired gluteal and thigh muscle groups, and overlapping tissue maps for subcutaneous fat, skeletal muscle, and torso fat (D'Antonoli et al., 2024).

Category	Count
Organs	20
Bones	18
Muscles	11
Vessels	7
Tissue types	3

This breadth is central to its identity. TotalSegmentator MRI is not primarily a specialist model for one body region or one MRI sequence; it is a generalist MRI segmenter designed to operate on real-world heterogeneity. At the same time, it remains narrower than CT TotalSegmentator in both anatomical completeness and class count (D'Antonoli et al., 2024).

2. Data sources, heterogeneity, and reference standard construction

The study assembled a total dataset of 525 images: 298 MR scans and 227 CT scans. The MRI cohort comprised 251 MR examinations randomly sampled from the University Hospital Basel PACS and 47 MR images from the Imaging Data Commons, while the CT cohort was drawn from the original TotalSegmentator dataset. After exclusions for missing slices or major artifacts, the final split was 495 images for training and 30 images for testing, with the internal test set containing only MR images (D'Antonoli et al., 2024).

A defining feature of the dataset is real-world diversity. The MRI data were sampled from routine clinical studies with random series selection within studies, and the paper emphasizes variation in age, pathology, scanner manufacturer, scanner model, acquisition site, body part, MRI sequence type, contrast weighting, echo time, repetition time, field strength, slice thickness, contrast agent usage, orientation, and resolution. The study reports 3 manufacturers, 4 sites, and 30 different scanners (D'Antonoli et al., 2024).

Reference annotations were not created by de novo manual tracing alone. Ground-truth labels were produced through a combination of manual segmentation, manual correction of existing model predictions, and an iterative learning workflow. Annotation was performed in the Nora Imaging Platform and supervised by a board-certified radiologist with 12 years of experience. The authors stress that all 298 MR scans ultimately had annotations that were manually reviewed and corrected if necessary (D'Antonoli et al., 2024).

This annotation design matters methodologically. It places TotalSegmentator MRI within the contemporary pattern of large-scale medical segmentation systems that rely on iterative expert correction of model-generated masks rather than one-pass manual annotation of every structure from scratch. A plausible implication is that the project prioritized label breadth and clinical heterogeneity over a narrowly controlled single-protocol reference dataset.

3. Model design and training strategy

The architecture is based on nnU-Net. The paper does not introduce a novel segmentation backbone; its innovations lie in dataset design, cross-modality training, annotation workflow, and the focus on broad MRI robustness. Two resolutions were trained: a main model at 1.5 mm isotropic resolution and a lower-resource model at 3 mm isotropic resolution (D'Antonoli et al., 2024).

Because a single 1.5 mm model could not practically handle all non-overlapping classes within typical workstation memory limits, the class set was partitioned. The paper states that 56 non-overlapping classes were divided into two separate models with 28 classes each, while the 3 overlapping tissue classes—skeletal muscle, subcutaneous fat, and torso fat—were trained in a separate model. A single-model 56-class variant was also tested and performed slightly worse than the split-model configuration (D'Antonoli et al., 2024).

Several implementation choices are explicitly reported. Normalization was set to “MR” for all experiments, even when training only on CT images. Segmentation mask resampling interpolation was set to order 0. Mirroring augmentation was removed because otherwise the model had difficulty distinguishing left from right structures. The authors also deliberately chose not to train and ensemble the standard 5 nnU-Net folds, in order to keep inference runtime lower (D'Antonoli et al., 2024).

CT inclusion is a core part of the training strategy. The main proposed model was trained on MR + CT together. The ablation study showed that, on the MR test set, the mixed MR + CT model performed better than an MR-only model, whereas on the CT test set the CT-only model performed slightly better than the mixed model. The paper interprets this as evidence that CT contributes additional anatomical supervision useful for MRI segmentation, even if MRI makes CT learning slightly harder when CT is evaluated alone (D'Antonoli et al., 2024).

4. Evaluation profile and empirical performance

On the internal MR test set, the main model achieved Dice = 0.824 [CI: 0.801, 0.842] and NSD = 0.882 [CI: 0.860, 0.900]. The lower-resource 3 mm model was substantially weaker, with Dice = 0.724 [CI: 0.683, 0.757] and NSD = 0.799 [CI: 0.760, 0.827] (D'Antonoli et al., 2024).

The paper includes two structured comparative evaluations. For the 40 structures supported by MRSegmentator, TotalSegmentator MRI obtained Dice 0.824 [CI: 0.796, 0.848] and NSD 0.889 [CI: 0.859, 0.913], versus Dice 0.762 [CI: 0.733, 0.787] and NSD 0.824 [CI: 0.792, 0.856] for MRSegmentator, with p < 0.001 for both Dice and NSD. For the 13 structures supported by the AMOS model, TotalSegmentator MRI obtained Dice 0.762 [CI: 0.692, 0.806] and NSD 0.826 [CI: 0.745, 0.871], versus Dice 0.542 [CI: 0.475, 0.606] and NSD 0.619 [CI: 0.557, 0.681] for AMOS, again with p < 0.001 (D'Antonoli et al., 2024).

External validation on the AMOS MR validation set yielded Dice 0.801 [CI: 0.780, 0.824] and NSD 0.883 [CI: 0.866, 0.897]. On that dataset, MRSegmentator scored Dice 0.836 [CI: 0.815, 0.856] and NSD 0.924 [CI: 0.910, 0.938], while the AMOS model scored Dice 0.907 [CI: 0.893, 0.919] and NSD 0.972 [CI: 0.965, 0.978]. The paper attributes the strong AMOS result on its own validation set to domain similarity between AMOS training and validation images (D'Antonoli et al., 2024).

A further result of practical significance is CT retention. On the original CT test set, TotalSegmentator MRI achieved Dice 0.960 [CI: 0.959, 0.962] and NSD 0.994 [CI: 0.993, 0.994], compared with Dice 0.970 [CI: 0.969, 0.971] and NSD 0.997 [CI: 0.996, 0.997] for the original CT-specialized TotalSegmentator. This indicates that mixed MRI-capable training preserved most of the original CT capability (D'Antonoli et al., 2024).

Performance is structure-dependent. Better-performing structures included larger, more consistently visible targets such as brain, lungs, liver, kidneys, vertebrae, larger muscles, bladder, and heart. Harder structures included adrenal glands, portal vein and splenic vein, iliac arteries, iliac veins, duodenum, and prostate in some settings. The paper attributes poor performance mainly to thick slices and anisotropy, low contrast outside the target region, small structure size, and ground-truth imperfections (D'Antonoli et al., 2024).

5. Position within the MRI segmentation ecosystem

The name “TotalSegmentator MRI” sits within a broader and partly overlapping MRI segmentation ecosystem. MRSegmentator is best understood as the current MRI-focused companion to TotalSegmentator rather than a literal one-to-one “MRI TotalSegmentator” with the same full label space; it segments 40 anatomical structures, uses nnU-Net, employs a mixed MRI+CT training strategy, and is not presented as an official 104-label MRI port of TotalSegmentator (Häntze et al., 2024). TotalVibeSegmentator, by contrast, is explicitly tied to VIBE torso MRI from cohorts such as NAKO and UK Biobank; it provides 71 semantic classes plus a separate 22-vertebra instance model and is best interpreted as a torso-focused MRI adaptation rather than a generic all-sequence MRI replacement for TotalSegmentator (Graf et al., 2024).

Benchmark studies clarify how these systems relate in practice. In a curated benchmark of 40 axial T1-weighted abdominal MRI volumes from the Duke Liver Dataset, TotalSegmentator MRI was evaluated as an off-the-shelf public model that segmented 59 structures and ranked second overall behind MRSegmentator: 77.7 ± 18.6 DSC and 10.8 ± 13.1 mm HD for TotalSegmentator MRI, versus 80.7 ± 18.6 DSC and 8.9 ± 10.4 mm HD for MRSegmentator and 74.3 ± 20.2 DSC and 16.4 ± 20.1 mm HD for TotalVibeSegmentator (Tran et al., 10 Apr 2025). A later benchmark on unseen public abdominal MRI datasets likewise concluded that MRSegmentator was most generalizable, while TotalSegmentator MRI remained competitive, more robust than MRISegmentator-Abdomen on unseen CHAOS sequences, and the fastest of the nnU-Net baselines at 39.60 ± 10.34 s with 31M trainable parameters (Krishnaswamy et al., 23 Jul 2025).

These results suggest a stable pattern. TotalSegmentator MRI is strongest as a broad, open, sequence-diverse MRI segmenter; MRSegmentator often leads on abdominal benchmarks; and TotalVibeSegmentator is most specifically aligned with VIBE torso epidemiology data. The systems are related, but they are not interchangeable.

6. Limitations, misconceptions, and practical interpretation

Several limitations are explicit in the TotalSegmentator MRI paper. Overall performance did not exceed Dice 0.9 because many MR images were lower resolution and lower quality than CT in routine practice. The MRI model currently segments fewer structures than CT TotalSegmentator. External validation was limited, memory usage was high enough that class splitting became necessary, the 3 mm lower-memory model lost substantial accuracy, and no explicit prospective clinical deployment study was reported. The authors identify three immediate future directions: add more structures, improve annotation quality, and expand the training dataset (D'Antonoli et al., 2024).

A recurrent misconception is that broad MRI coverage implies universal adequacy for every organ-specific workflow. Evidence from a dedicated prostate mpMRI study argues against that reading: a prostate-specific multimodal nnU-Net v2 model achieved Dice 0.82 on external validation, whereas TotalSegmentator achieved Dice 0.15, with failure characterized primarily by severe under-segmentation of the gland (Rodriguez-Belenguer et al., 2 Apr 2026). This suggests that sequence-independent, multi-structure MRI segmentation does not eliminate the need for task-specific multimodal models when the target is a small, clinically specialized structure.

From a deployment perspective, TotalSegmentator MRI is explicitly presented as open-source and easy-to-use. The paper states that the model, training dataset, and annotations are public; it also points to a ready-to-use online tool at totalsegmentator.com and the TotalSegmentator toolkit on GitHub (D'Antonoli et al., 2024). Its practical role is therefore clear: it is a general-purpose MRI counterpart within the TotalSegmentator ecosystem, designed for broad anatomical segmentation across heterogeneous routine MR images, but not a finished solution to every MRI segmentation problem or a strict MRI replica of the full CT TotalSegmentator label universe.