MRI segmentation models can now label 80 structures, but deployment is the hard part

Radiology teams are getting something they have been asking for since 2023: multi-organ, sequence-independent MRI segmentation that is usable outside a research sandbox. The newest open-source models are moving from “one organ, one model” to “one model, many structures,” and the operational question is quickly changing from “is it accurate?” to “who trusts it, where, and how is it governed?”

Two separate research tracks show the shift. TotalSegmentator MRI, built on nnU-Net, is reported to segment 80 anatomical structures across MRI sequences, according to Radiology Business. In parallel, an arXiv preprint describing MRSegmentator reports consistent multi-modality segmentation across 40 classes in MRI and CT, with wide variation by structure. Together they give imaging operators a more realistic benchmark for 2026 planning: segmentation is becoming cheap and fast, but it still demands deliberate QA, workflow integration, and data controls.

TotalSegmentator MRI pushes the “one model, many structures” idea into MRI

On Feb. 18, 2025, Radiology Business reported on TotalSegmentator MRI, an open-source nnU-Net model designed to segment major organs across MRI sequences. The reported scope, 80 anatomical structures, matters because it maps to the messy reality of enterprise MRI, where protocol variability is the rule and “sequence-specific” tools often fail the moment they leave a controlled dataset.

AuntMinnie, also covering the Radiology paper that day, described the training and test design. The team trained the model using a dataset of 616 MR and 527 CT exams, and evaluated performance on an internal test set of 55 MRI exams.

Radiology Business described TotalSegmentator MRI as an open-source nnU-Net model that provides rapid segmentation of all major organs across all MRI sequences and could reduce the time providers spend completing manual segmentations.

MRSegmentator’s per-organ scores show why averages don’t run a clinic

The MRSegmentator preprint on arXiv (v2 posted May 13, 2024) provides a different kind of operationally useful detail: performance dispersion. The authors report training on 1,200 manually annotated MRI scans from UK Biobank, plus 221 in-house MRI scans and 1,228 CT scans, using cross-modality transfer learning from CT segmentation models and a human-in-the-loop annotation workflow.

On headline organs, the reported Dice Similarity Coefficient (DSC) is high: 0.97 for right and left lungs and 0.95 for the heart. Liver (0.96) and kidneys (0.95 left and right) were also strong. But the preprint reports materially lower DSC for smaller or more complex targets, including 0.54 for portal and splenic veins and about 0.61 to 0.65 for adrenal glands.

For enterprise operators, that spread is the point. A single “mean Dice” is fine for academic comparison, but it is not an operating policy. If a model is excellent on large organs yet unstable on smaller vessels, the hospital needs an explicit workflow decision: which structures can auto-populate a report or a radiation contouring workspace, which require mandatory review, and which should be hidden unless confidence is high. The arXiv paper does not describe whether low-confidence segmentations are flagged for manual review, leaving that as an integration responsibility for the deploying organization.

Where this lands first: oncology contouring and neuro volumetry

Segmentation is not a standalone deliverable in most care settings. It is an upstream input to something that gets audited: a radiotherapy plan, a volumetry report, or a longitudinal disease monitoring program.

Health Imaging reported in June 2025 on iSeg, a lung tumor segmentation tool developed at Northwestern Medicine, positioned for radiation therapy planning. The reporting described iSeg as matching physician inter-observer variability. Health Imaging did not report that missed areas were linked to worse outcomes, so that claim is removed. That is not the same as proof that AI improves outcomes, but it does sharpen the operational value proposition for tumor boards and radiation oncology leaders focused on consistency.

In neurology workflows, the operational lever is often speed and throughput. A 2023 paper in Scientific Reports (Nature) compared deep learning approaches with FreeSurfer for segmenting six brain structures used in Parkinsonian syndrome evaluation. The authors reported segmentation times of about 51 seconds per patient for a CNN model and about 1,102 seconds for a vision transformer model, versus about 15,735 seconds for FreeSurfer. The paper also reported Dice scores above 0.85 and AUCs above 0.8 for several classification tasks, suggesting that faster pipelines can be viable without sacrificing diagnostic performance.

The enterprise bottleneck moves to governance, privacy, and responsibility

Once segmentation becomes “good enough” for many structures, the hard work shifts to who is responsible for the output and how it is monitored.

Those constraints show up in mundane places: the DICOM routing rules that decide which series even reaches an AI server; the identity and access management policies that determine whether a research model can touch clinical images; the audit trail needed when a contour becomes part of a radiation therapy plan; and the clinical policy that defines when a radiologist must review an auto-segmentation versus when a technologist can accept it.

A broader survey article in Bioengineering (Basel), available via PubMed Central, frames the same operational reality: modern imaging AI improvements are increasingly driven by deep learning architectures, but deployment depends on integration into diagnostic and care processes, not just algorithmic performance. That’s old news academically, but it is still where many health systems get stuck.

Where to pressure-test these models in 2026 deployments

For imaging leaders and oncology operations teams, the shared message across these sources is that “open-source” does not mean “plug-and-play,” and “high Dice” does not mean “low risk.” The models are increasingly capable. The implementation is where projects succeed or stall.

• Define per-structure acceptance rules before any pilot: which organs can auto-fill measurements, which are “draft only,” and which are excluded unless manually verified.
• Ask for reporting that matches how the tool will be used. According to AuntMinnie, TotalSegmentator MRI was trained on 616 MR and 527 CT exams and targets 80 anatomic structures. Teams still need clarity on where performance varies by use case.
• Decide who owns QA and incident response. Radiology Business reported that iSeg is intended to improve tumor segmentation prior to radiation therapy, a setting where escalation paths for suspect contours should be defined in the workflow.
• Treat data governance as a first-class requirement: image access, de-identification where needed, retention, and audit logs should be specified alongside performance.