mrsegmentator 2.0.0

Multi-Modality Segmentation of 40+10 Classes in MRI and CT

MRSegmentator: Multi-Modality Segmentation of 40+10 Classes in MRI and CT

---

Detect and segment 40 classes in MRI and CT of the abdominal / pelvic / thorax region. Now with 10 additional body composition classes in MRI

Contrary to CT scans, where tools for automatic multi-structure segmentation are quite mature, segmentation tasks in MRI scans are often either focused on the brain region or on a subset of few organs in other body regions. MRSegmentator aims to extend this and accurately segment 40 organs and structures in human MRI scans of the abdominal, pelvic and thorax regions. The segmentation works well on different sequence types, including T1- and T2-weighted, Dixon sequences and even CT images.

Updates (v2.0.0)

• We added 10 additional body composition classes, exclusively for MRI. You can segment them by setting --body_comp
• You can accelerate segmentation with the --fast flag

Installation

Install MRSegmentator with pip:

# Create virtual environment
conda create -n mrseg python=3.11 pip
conda activate mrseg

# Install MRSegmentator
python -m pip install mrsegmentator

If the installed pytorch version is not compatible to your system, you might need to install it manually. Please refer to PyTorch. MRSegmentator requires torch <= 2.3.1.

Docker Image

You can run an MRSegmentator (v1.2) Docker image directly from MHub.

$input_dir=/path/to/input
$output_dir=/path/to/output

docker run --rm -t --gpus all --network=none -v $input_dir:/app/data/input_data:ro -v $output_dir:/app/data/output_data mhubai/mrsegmentator:latest --workflow default

Inference

MRSegmentator segments all .nii/.nii.gz/.mha/.nrrd files in an input directory and writes segmentations to the specified output directory. To speed up segmentation you can increase the --batchsize or select a single model for inference with --fast. MRSegmentator requires a lot of memory and can run into OutOfMemory exceptions when used on very large images. You can reduce memory usage by setting --split_level to 1 or 2. Be aware that this increases runtime. Read more about the options in the Evaluation section.

You can now also run MRSegmentator on DICOM directories, in which case it produces a DICOM SEG. (Make sure that there is only a single series UID in the directory). You can also convert previously created segmentations back to DICOM SEG (see dcm_helper).

mrsegmentator --input <file / directory / DICOM directory>

Options:

-i, --input <str> [required] # input directory or file

-o --outdir <str>   # output directory
--body_comp         # segment 10 body composition classes in MRI
--fast              # accelerate segmentation by disabling ensembling, mirroring, and by using a larget step size
--postfix <str>     # postfix that will be added to segmentations, default: "seg"
--cpu_only          # don't use a gpu

# memory (mutually exclusive)
--batchsize <int>   # number of images that can be loaded to memory at the same time, default: 8 
--split_level <int> # split images to reduce memory usage. Images are split recursively: A split level of x will produce 2^x smaller images

# debugging
--log_level <["DEBUG", "INFO", "WARNING", "ERROR"]> # Default: INFO
--no_tqdm               # disable tqdm progress bars

# experimental
--split_margin <int>    # split images with an overlap of 2xmargin to avoid hard cutt-offs between segmentations of top and bottom image, default: 3
--nproc <int>           # number of processes
--nproc_export <int>    # number of processes for exporting the segmentations
--fold <int>            # specify a single fold for segmentation

Python API

from mrsegmentator import inference
import os

outdir = "outputdir"
images = [f.path for f in os.scandir("image_dir")]

inference.infer(images, outdir, model_name="base")

Manage Weights

Weights are automatically downloaded to ~/.mrsegmentator. If you want to specify a custom weights directory you can do so by setting an environmental variable: export MRSEG_WEIGHTS_PATH=<path>.

You can also use MRSegmentator as an interface to run other nnunetv2 models by pointing the MRSEG_WEIGHTS_PATH variable directly to the nnunetv2 folder with the plans.json file. By doing so you have a nice inference interface with all the included pre-processing:

• force LPS orientation during loading
• support for various file types
• fast mode (--fast) and memory efficiency (--split_level 2)

How To Cite

If you use our work in your research, please cite our article: https://doi.org/10.1148/ryai.240777.

If you use the body composition classes, please additionally cite this preprint: https://doi.org/10.1101/2025.06.03.25328867

Class details

Index	Class
0	background
1	spleen
2	right_kidney
3	left_kidney
4	gallbladder
5	liver
6	stomach
7	pancreas
8	right_adrenal_gland
9	left_adrenal_gland
10	left_lung
11	right_lung
12	heart
13	aorta
14	inferior_vena_cava
15	portal_vein_and_splenic_vein
16	left_iliac_artery
17	right_iliac_artery
18	left_iliac_vena
19	right_iliac_vena
20	esophagus
21	small_bowel
22	duodenum
23	colon
24	urinary_bladder
25	spine
26	sacrum
27	left_hip
28	right_hip
29	left_femur
30	right_femur
31	left_autochthonous_muscle
32	right_autochthonous_muscle
33	left_iliopsoas_muscle
34	right_iliopsoas_muscle
35	left_gluteus_maximus
36	right_gluteus_maximus
37	left_gluteus_medius
38	right_gluteus_medius
39	left_gluteus_minimus
40	right_gluteus_minimus

Body Composition (Exclusively MRI)

Index	Class
1	subcutaneous_fat
2	visceral_fat
3	left_rectus_abdominis
4	right_rectus_abdominis
5	left_oblique_muscle
6	right_oblique_muscle
7	left_quadratus_lumborum
8	right_quadratus_lumborum
9	abdominal_subcutaneous_fat
10	gluteofemoral_fat

Testing

The test suite has two modes:

Smoke tests — no model weights or real images required, runs in seconds

Integration tests — requires model weights and at least one real MRI/CT image

make smoke

cp tests/mrsegmentator/integration_config.example.py tests/mrsegmentator/integration_config.py
#fill in MODEL_PATH and TEST_IMAGES
make full

Integration tests run inference exactly once and validate that outputs have correct labels, geometry, and are readable. Slice figures (axial / coronal / sagittal) are saved to reports/figures/ after each run.

Acknowledgements

This work was in large parts funded by the Wilhelm Sander Foundation. Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or European Health and Digital Executive Agency (HADEA). Neither the European Union nor the granting authority can be held responsible for them.

Changelog

v2.0.0 (06/07/2027)

Feature

• 10 additional classes for body composition in MRI
• weights management supports muliple weights
• dedicatated weights directory as ~/.mrsegmentator
• env variable MRSEG_WEIGHTS_PATH allows to link to a custom weights dir
• if MRSEG_WEIGHTS_PATH points to any nnUNet results direcotry, these weights can be loaded as well
• More CI tests, with full integration workflows
• More GitHub action workflows
• Updated supported python versions to 3.10-3-13. (3.9 is not working)
• Added fast inference mode with --fast

v1.3.2 (26/05/2026)

Fix

• Fixed incorrect handling of image orientations such as PIL or ASL

v1.3.1 (11/08/2025)

Fix

• fixed incorrect file names that occured when mapping DICOM to NIfTI

v1.3.0 (09/08/2025)

Feature

• Automatically run on CPU if no GPU was detected
• Add support for .mha and .nrrd
• Add support for DICOM and DICOM SEG
• Enable better logging:
  • control log lebel (DEBUG, INFO, WARNING)
  • add flag --no_tqdm to disable progress bars
  • remove --verbose flag (replaced by --log_level DEBUG)

v1.2.3 (05/02/2025)

Feature

• Print image and subvolume size if splitting is used
• Add option --split_margin to allow to change overlap between splitted volumes

Fix

• Set pytotch version to <= 2.3.1
• Set python version to < 3.13
• ==> Fixes toch.pickle error due to updated dependency- Supress torch.load future warning, introduced by nnunet
• Increase default split_margin from 2 to 3

v1.2.2 (11/12/2024)

Feature

• Print segmentation time after finishing

Fix

• Supress torch.load future warning, introduced by nnunet
• Print version number of custom weight directories, if they are specified

v1.2.0 (22/08/2024)

Feature

• Add NAKO data to training pipeline
• Update weights

Fix

• Make ensemble prediction default for Python API

---

v1.1.2 (24/06/2024)

Fix

• Change python_requires from 3.11 to 3.9
• Remove monai dependency

v1.1.0 (18/05/2024)

Feature

• Update model weights with weights trained by nnUNetTrainerNoMirroring

Fix

• Remove postprocessing remap_left_right(...). It is not needed anymore.

---

v1.0.0 (10/05/2024)

• First release of MRSegmentator