Angelou0516/TotalSegmentatorMR

TotalSegmentator MRI (v2.0.0)

Public release of the TotalSegmentator MRI dataset: 616 heterogeneous MRI scans (T1, T2, PD, DIXON-style and other sequences, multiple field strengths, scanners, slice thicknesses, contrast agents) with voxel-wise annotations of 50 anatomical structures spanning the whole body.

Dataset Summary

Field	Details
Modality	MRI (sequence-independent: T1w / T2w / PD / DIXON / mixed)
Body Part	Whole-body / multi-organ
Structures	50 anatomical regions per scan
Subjects	616 MRI scans (561 train / 55 internal test, per meta.csv)
Sources	University Hospital Basel PACS (2011–2023) + Imaging Data Commons
Total Size	~5.8 GB
License	CC BY-NC-SA 2.0 (non-commercial)
Source	Zenodo: https://zenodo.org/records/14710732

The dataset is the MR portion of the study cohort described in the Radiology paper. External AMOS and CHAOS test sets referenced by the paper are not redistributed here — they live in their own datasets and are referenced by meta.csv only.

Annotation Pipeline

Ground-truth masks were produced iteratively:

1. Initial segmentations from public CT/MR models where available.
2. Manual segmentation of the first 10 patients to bootstrap an internal nnU-Net.
3. Predicted masks were manually refined; the model was retrained after every additional 125 reviewed patients.
4. All annotations were manually reviewed and corrected when needed by a board-certified radiologist with 12 years of experience, blinded to radiology reports.

A single GT mask volume per scan is provided. There is no "predicted vs. corrected" duality.

Data Structure

TotalSegmentatorMR/
├── meta.csv                                        # split + scanner metadata (semicolon-separated)
└── s{0001..0616}/
    ├── mri.nii.gz                                  # MR scan
    └── segmentations/
        ├── adrenal_gland_left.nii.gz
        ├── adrenal_gland_right.nii.gz
        ├── aorta.nii.gz
        ├── ...                                     # one binary mask per anatomical structure
        └── (50 files total)

Each segmentations/<structure>.nii.gz is a binary mask aligned to the corresponding mri.nii.gz.

meta.csv columns

image_id; split; age; gender; institute; study_type; manufacturer; scanner_model; slice_thickness; scanning_sequence; repetition_time; echo_time; magnetic_field_strength; source

The split column encodes train/test partitioning; source distinguishes Basel PACS vs IDC provenance.

Versioning

• v2.0.0 (current, this upload) — 616 scans, 50 anatomical regions. subcutaneous_fat, torso_fat, skeletal_muscle, and face_region were moved into separate subtask models (tissue_types_mr, face_mr).
• v1.0.0 — 298 scans, 56 anatomical regions.

Citation

@article{akinci2025totalsegmentatormr,
  title   = {{TotalSegmentator MRI}: Robust Sequence-independent Segmentation of Multiple Anatomic Structures in {MRI}},
  author  = {Akinci D'Antonoli, Tugba and Berger, Lucas K. and Indrakanti, Ashraya K. and Vishwanathan, Nathan and Wei{\ss}, Jakob and Jung, Matthias and Berkarda, Zeynep and Rau, Alexander and Reisert, Marco and K{\"u}stner, Thomas and Walter, Annette and Merkle, Elmar M. and Boll, Daniel T. and Breit, Hanns-Christian and Nicoli, Andrea P. and Segeroth, Martin and Cyriac, Joshy and Yang, Shan and Wasserthal, Jakob},
  journal = {Radiology},
  year    = {2025},
  doi     = {10.1148/radiol.241613}
}

@article{wasserthal2023totalsegmentator,
  title   = {{TotalSegmentator}: Robust Segmentation of 104 Anatomic Structures in {CT} Images},
  author  = {Wasserthal, Jakob and Breit, Hanns-Christian and Meyer, Manfred T. and Pradella, Maurice and Hinck, Daniel and Sauter, Alexander W. and Heye, Tobias and Boll, Daniel T. and Cyriac, Joshy and Yang, Shan and Bach, Michael and Segeroth, Martin},
  journal = {Radiology: Artificial Intelligence},
  year    = {2023},
  doi     = {10.1148/ryai.230024}
}