MRISegmentator-Abdomen: Automated MRI Segmentation

• The paper introduces a fully automated 3D nnU-Net model that segments 62 abdominal organs using a voxel-level annotated, multi-phase T1-weighted MRI dataset.
• The methodology employs a two-stage annotation pipeline with synthetic MR-to-CT translation and iterative expert refinement, achieving high Dice and NSD scores across diverse datasets.
• MRISegmentator-Abdomen supports clinical applications such as radiotherapy planning and body composition analysis by providing a comprehensive segmentation benchmark for abdominal MRI.

Searching arXiv for MRISegmentator-Abdomen and closely related abdominal MRI segmentation/benchmark papers. MRISegmentator-Abdomen is a fully automated multi-organ and structure segmentation tool for T1-weighted abdominal MRI that was introduced to delineate 62 abdominal organs and structures from routine MRI acquisitions (Zhuang et al., 2024). It was developed alongside a voxel-level annotated dataset of 780 T1-weighted abdominal MRI series from 195 patients, with four axial phases per patient—pre-contrast, arterial, venous, and delayed—and was evaluated on an internal held-out test set as well as the external AMOS22 and Duke Liver datasets (Zhuang et al., 2024). In the originating report, the method is implemented as a 3D nnU-Net model and is referred to as “MRISegmentator” in short (Zhuang et al., 2024). Subsequent benchmarking positioned it among the state-of-the-art open-source abdominal MRI segmentation tools, while also highlighting sequence-dependent generalization limits outside its training distribution (Krishnaswamy et al., 23 Jul 2025).

1. Definition and scope

MRISegmentator-Abdomen addresses automated segmentation of abdominal MRI at a granularity substantially broader than earlier abdominal MRI approaches that focused on a limited number of structures (Zhuang et al., 2024). The reported target set comprises 62 abdominal organs and structures, and the accompanying dataset contains voxel-level annotations for each of those targets across all four T1-weighted contrast phases collected for each patient (Zhuang et al., 2024).

The training cohort was assembled from 195 patients imaged at the National Institutes of Health Clinical Center between January 2019 and October 2021, with same-day abdominal MRI and CT available for each patient (Zhuang et al., 2024). For each patient, four axial T1-weighted MRI series were collected—pre-contrast, arterial, portal venous, and delayed—yielding 780 volumes and 69,248 individual slices (Zhuang et al., 2024). The MRI spatial resolution varied within in-plane pixel spacing of 0.94–1.47 mm and slice thickness of 3.0–3.3 mm, and both 1.5 T and 3 T scanners were used (Zhuang et al., 2024).

The motivation given for the tool is the lack, at the time, of a publicly available abdominal MRI dataset with voxel-level annotations of multiple organs and structures, and correspondingly the absence of a publicly available multi-structure segmentation tool for abdominal MRI (Zhuang et al., 2024). A plausible implication is that MRISegmentator-Abdomen was designed not only as a model but also as an annotation and benchmarking resource intended to close a dataset bottleneck in abdominal MRI segmentation.

2. Dataset construction and annotation workflow

The annotation strategy was explicitly semi-automated and proceeded in two stages (Zhuang et al., 2024). In the first, “cross-domain” stage, each MRI was translated synthetically into CT contrast using a deep MR-to-CT network; the synthetic CT was then segmented with TotalSegmentator to produce pseudo-annotations for 15 major organs and 7 large vessels (Zhuang et al., 2024). In parallel, the patient’s true CT series, acquired the same day, were deformably registered to the MRI, segmented by TotalSegmentator to yield masks for 8 muscle groups and 32 osseous structures, and then manually refined by an expert (Zhuang et al., 2024).

In the second, “iterative learning” stage, the initial pseudo-labels were corrected and augmented over four rounds, denoted R1–R4, of “train-refine-retrain” (Zhuang et al., 2024). At each round, an nnU-Net model was trained on the current annotations, deployed to predict on unlabeled volumes, and its predictions were refined by a radiologist (Zhuang et al., 2024). The final result was fully corrected voxel-wise maps for all 62 structures in every volume (Zhuang et al., 2024).

The patient-level split used 135 patients and 540 volumes for training and 60 patients and 240 volumes for a held-out internal test set (Zhuang et al., 2024). All annotations in the test set were reviewed by both a radiology resident and a board-certified radiologist prior to final correction (Zhuang et al., 2024). This review protocol is notable because it constrains label quality on the evaluation set more tightly than would be expected from a purely single-rater workflow. A plausible implication is that the reported internal performance reflects both algorithmic behavior and a relatively stringent reference standard.

A later benchmark summarized the same training regime somewhat differently but consistently in substance: it reported 780 T1-weighted MRI volumes from a proprietary in-house cohort, 62 annotated anatomical structures, and labels transferred from CT segmentations via a learned MR-to-pseudo-CT translation plus iterative expert refinement using MONAI Label (Krishnaswamy et al., 23 Jul 2025). That benchmark also noted the presence of pathologies such as liver tumors, pancreatic cysts, and assorted abdominal abnormalities in the training data (Krishnaswamy et al., 23 Jul 2025).

3. Model architecture, optimization, and inference

MRISegmentator-Abdomen is implemented using the 3D full-resolution instantiation of nnU-Net (Zhuang et al., 2024). In the originating description, the network is automatically configured as a U-shaped encoder-decoder with five resolution stages, with each stage consisting of two blocks of convolution, instance normalization, and leaky ReLU activation, and with down- and up-sampling by stride-2 operations (Zhuang et al., 2024). The training objective is the sum of Dice loss and binary cross-entropy loss,

$L = DSC_{\text{loss}} + BCE_{\text{loss}},$

optimized with Adam (Zhuang et al., 2024).

The reported training schedule used 2,000 epochs, an initial learning rate of $10^{-2}$ decayed by a polynomial schedule, a batch size of 2 volumes, and five-fold cross-validation (Zhuang et al., 2024). At inference, the five models were ensembled by soft voting (Zhuang et al., 2024). All experiments ran on an NVIDIA A100 80 GB GPU in the NIH Biowulf cluster (Zhuang et al., 2024). Input volumes were z-score normalized per channel and cropped to nonzero bounding boxes under the nnU-Net pipeline, which also supplied on-the-fly 3D augmentations including random affine deformations, elastic deformations, intensity scaling or noise, gamma correction, and mirroring (Zhuang et al., 2024).

A subsequent benchmark characterized the same system as a standard 3D U-Net configured by nnU-Net with four down-sampling and up-sampling levels, two successive $3 \times 3 \times 3$ convolutions per level, instance normalization, LeakyReLU, a final $1 \times 1 \times 1$ convolution to $C=62$ output channels, and softmax, with no additional attention or transformer modules reported (Krishnaswamy et al., 23 Jul 2025). That benchmark also described nnU-Net-style preprocessing—resampling to the median voxel spacing, clipping intensities at the 0.5 and 99.5 percentiles, z-score normalization, and automatic body cropping—and patch-based optimization with patch size $48 \times 160 \times 192$, batch size 2, stochastic gradient descent with Nesterov momentum 0.99, weight decay $3 \times 10^{-5}$, and a polynomial learning-rate decay over approximately 1,000 epochs (Krishnaswamy et al., 23 Jul 2025). For inference, it reported sliding-window prediction with one-quarter overlap and post-processing by removal of isolated components smaller than 10,000 voxels for each label (Krishnaswamy et al., 23 Jul 2025).

These two descriptions are not identical in implementation detail. The common denominator is that MRISegmentator-Abdomen is framed as an nnU-Net-based 3D multi-class segmentation system whose novelty lies primarily in the training data and annotation pipeline rather than in a bespoke architectural departure (Zhuang et al., 2024, Krishnaswamy et al., 23 Jul 2025).

4. Segmentation targets and quantitative performance

The primary evaluation metrics in the originating study were the Dice Similarity Coefficient (DSC) and the Normalized Surface Distance (NSD) (Zhuang et al., 2024). On the internal test set, comprising 240 volumes and all 62 structures, MRISegmentator achieved a mean DSC of $0.861 \pm 0.170$ and a mean NSD of $0.924 \pm 0.163$ (Zhuang et al., 2024).

Performance was further stratified into four structure groups (Zhuang et al., 2024):

Group	Structures	Performance
G1	15 major organs	DSC $= 0.918 \pm 0.069$, NSD $10^{-2}$0
G2	7 vessels	DSC $10^{-2}$1, NSD $10^{-2}$2
G3	8 muscles	DSC $10^{-2}$3, NSD $10^{-2}$4
G4	32 bones	DSC $10^{-2}$5, NSD $10^{-2}$6

Phase-stratified internal performance was reported as follows (Zhuang et al., 2024):

Phase	DSC	NSD
PRE	$10^{-2}$7	$10^{-2}$8
ART	$10^{-2}$9	$3 \times 3 \times 3$0
VEN	$3 \times 3 \times 3$1	$3 \times 3 \times 3$2
DEL	$3 \times 3 \times 3$3	$3 \times 3 \times 3$4

A mixed-effects model confirmed that post-contrast phases significantly outperformed PRE with $3 \times 3 \times 3$5, while ART, VEN, and DEL did not differ significantly from one another (Zhuang et al., 2024). This suggests that contrast enhancement systematically improves separability for the target classes in this dataset, but that the three post-contrast phases provide broadly similar segmentation conditions.

Per-structure rankings followed expected size and contrast trends (Zhuang et al., 2024). Large, high-contrast organs such as the liver, spleen, kidneys, and back muscles were among the highest-performing classes, with reported DSCs including liver $3 \times 3 \times 3$6, spleen $3 \times 3 \times 3$7, right kidney $3 \times 3 \times 3$8, and left kidney $3 \times 3 \times 3$9 (Zhuang et al., 2024). The smallest or morphologically thin structures were more challenging: adrenal glands were approximately 0.66 Dice, upper ribs fell to approximately 0.48, and the transitional vertebra T7 was 0.682 (Zhuang et al., 2024).

External validation showed mean DSC and NSD of $1 \times 1 \times 1$0 and $1 \times 1 \times 1$1 on AMOS22, and mean DSC and NSD of $1 \times 1 \times 1$2 and $1 \times 1 \times 1$3 on Duke Liver (Zhuang et al., 2024). On the 20-volume AMOS22 validation subset, the model, trained only on its internal data, achieved DSC $1 \times 1 \times 1$4 and ranked second among methods trained directly on AMOS22 (Zhuang et al., 2024). On the Duke Liver set, PRE, ART, and VEN phases yielded liver DSCs of 0.912, 0.946, and 0.942 and NSDs of 0.903, 0.929, and 0.956, respectively (Zhuang et al., 2024).

5. Generalization, robustness, and benchmark positioning

MRISegmentator-Abdomen was later benchmarked against MRSegmentator, TotalSegmentator MRI, and ABDSynth on public datasets not seen during training (Krishnaswamy et al., 23 Jul 2025). In that study, it was described as one of three state-of-the-art abdominal MRI segmentation tools, but its generalization profile was more sequence-sensitive than that of MRSegmentator (Krishnaswamy et al., 23 Jul 2025).

On AMOS, the benchmark reported strong Dice values for major organs, including liver $1 \times 1 \times 1$5, spleen $1 \times 1 \times 1$6, left kidney $1 \times 1 \times 1$7, right kidney $1 \times 1 \times 1$8, pancreas $1 \times 1 \times 1$9, and stomach $C=62$0 (Krishnaswamy et al., 23 Jul 2025). However, it also reported very large HD95 outliers, such as liver $C=62$1 mm and right kidney $C=62$2 mm, which it interpreted as evidence of fragmented or anatomically implausible predictions in some volumes (Krishnaswamy et al., 23 Jul 2025).

On the CHAOS dataset, which contained sequences absent from the model’s training set, the benchmark reported marked performance degradation (Krishnaswamy et al., 23 Jul 2025). For T1 dual in-phase, Dice values were 0.81 for liver, 0.34 for spleen, 0.63 for left kidney, and 0.78 for right kidney; for T1 dual out-phase, 0.83, 0.45, 0.77, and 0.74; and for T2 SPIR, 0.88, 0.87, 0.92, and 0.91, respectively (Krishnaswamy et al., 23 Jul 2025). The same benchmark concluded that MRISegmentator-Abdomen performed well on AMOS T1 sequences but lacked cross-sequence robustness on unseen T1 dual in/out-phase acquisitions and showed high HD95 values on several settings (Krishnaswamy et al., 23 Jul 2025).

For LiverHCCSeg, the benchmark reported liver Dice $C=62$3 and HD95 $C=62$4 mm against one rater (Krishnaswamy et al., 23 Jul 2025). It noted that this Dice approached inter-rater agreement, although HD95 remained elevated compared with MRSegmentator (Krishnaswamy et al., 23 Jul 2025).

The benchmark’s comparative interpretation was that MRSegmentator achieved the best performance and was most generalizable overall (Krishnaswamy et al., 23 Jul 2025). In contrast, MRISegmentator-Abdomen was portrayed as highly accurate on T1 acquisitions similar to its training data but less robust to commonly used MRI protocols outside that distribution (Krishnaswamy et al., 23 Jul 2025). This suggests that the model’s dense annotation scope and strong in-domain performance do not automatically translate into protocol-agnostic behavior.

6. Failure modes, deployment characteristics, and applications

The originating study identified several failure patterns concentrated in small, thin, and low-contrast targets, as well as in orientations or sequences not represented in training (Zhuang et al., 2024). Specifically, adrenal glands, duodenum, thin ribs, boundary regions with abrupt tissue transitions, and coronal T1-weighted volumes including chest structures were reported as challenging (Zhuang et al., 2024). When the model encountered coronal or sagittal volumes incorporating portions of the lower chest, lung parenchyma and upper ribs were occasionally mis-segmented as bony structures (Zhuang et al., 2024). The study also stated that the tool was robust to a variety of pathologies, including liver metastases and renal lesions, and contrasted this with a comparator nnUNet-AMOS model trained only on “healthy” AMOS22 scans, which often failed to adhere to lesion boundaries in tumorous cases (Zhuang et al., 2024).

The benchmark paper emphasized somewhat different failure characteristics: sequence sensitivity, spatial outliers reflected by large HD95, and the operational cost of the annotation pipeline (Krishnaswamy et al., 23 Jul 2025). It identified spurious island-like predictions, missed anatomy, and sensitivity to training-sequence scope as salient limitations, and it characterized the cross-modality pseudo-CT plus iterative manual refinement workflow on nearly 800 MRI scans as time- and expert-intensive (Krishnaswamy et al., 23 Jul 2025).

For deployment, the model is used through the nnU-Net inference pipeline, accepts NIfTI inputs of any one of the four T1-weighted phases, performs its own normalization and cropping, and outputs a multi-label NIfTI segmentation (Zhuang et al., 2024). Dependencies include Python $C=62$5, PyTorch, nnU-Net, and standard Python libraries, and a Docker container is provided (Zhuang et al., 2024). Reported runtime for a single 80–96 slice volume is approximately 1–2 minutes on a single GPU such as an A100, or roughly 5–10 minutes on a high-end CPU (Zhuang et al., 2024).

The intended application areas named in the study include radiotherapy planning, opportunistic body composition analysis, longitudinal disease monitoring, and radiomics and AI research (Zhuang et al., 2024). Because the segmenter covers 62 organs and structures rather than a narrow organ subset, a plausible implication is that it is particularly suited to retrospective cohort analysis and multi-task downstream pipelines in which anatomical masks are reused for several quantitative tasks.

7. Relation to adjacent abdominal MRI segmentation methods

MRISegmentator-Abdomen emerged within a rapidly developing ecosystem of abdominal MRI segmentation tools. Earlier MRI-focused systems generally addressed fewer targets. ALAMO, for example, was a fully automated abdominal MR segmentation framework on 3 T T1-VIBE images for 10 organs-at-risk, using a 2D U-Net with dense connections, deep supervision, and multi-view aggregation, and reported overall mean DSC $C=62$6 on its 20-subject test set (Chen et al., 2019). ProRSeg targeted gastrointestinal organs at risk on MRI and CBCT, using joint registration-segmentation and reporting MRI DSC values of $C=62$7 for liver, $C=62$8 for large bowel, $C=62$9 for small bowel, and $48 \times 160 \times 192$0 for stomach-duodenum (Jiang et al., 2022). These methods addressed narrower label sets than MRISegmentator-Abdomen.

Among contemporary open-source tools, MRSegmentator and TotalVibeSegmentator provide relevant comparators. MRSegmentator segments 40 classes in MRI and CT, is also built on nnU-Net, and was trained on 1,200 UK Biobank MRI scans, 221 in-house MRI scans, and 1,228 CT scans with a human-in-the-loop annotation workflow (Häntze et al., 2024). In a benchmark on multi-parametric T1-weighted abdominal MRI, MRSegmentator achieved a pooled Dice score of $48 \times 160 \times 192$1 and Hausdorff Distance error of $48 \times 160 \times 192$2 mm on 40 curated Duke Liver volumes, outperforming TotalSegmentator MRI and TotalVibeSegmentator across sequence types with $48 \times 160 \times 192$3 (Tran et al., 10 Apr 2025). TotalVibeSegmentator, by contrast, focuses on VIBE images, produces 71 semantic torso labels, and reported an average Dice score of $48 \times 160 \times 192$4 overall 71 labels, with Dice-score above 0.90 on abdominal organs except for the pancreas at 0.70 (Graf et al., 2024).

Within this comparative landscape, MRISegmentator-Abdomen is distinguished primarily by its 62-class T1-weighted abdominal MRI scope and its paired dataset-and-tool release model (Zhuang et al., 2024). The later benchmark’s conclusion that MRSegmentator was more generalizable does not negate MRISegmentator-Abdomen’s importance; rather, it indicates a trade-off between dense in-domain annotation and broader sequence coverage in training (Krishnaswamy et al., 23 Jul 2025). This suggests that MRISegmentator-Abdomen is especially strong as a resource and tool for T1-weighted abdominal MRI resembling its NIH acquisition regime, while broader protocol robustness remains an open area for further work.