mrsegmentator 1.3.2

Multi-Modality Segmentation of 40 Classes in MRI and CT

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

---

Detect and segment 40 classes in MRI and CT of the abdominal / pelvic / thorax region

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 (v1.3.0)

• Our paper has been published! Read more about MRSegmentator in Radiology AI: https://doi.org/10.1148/ryai.240777
• We support .mha and .nrrd files now
• We support DICOM now: If a DICOM directory is used as input a corresponding DICOM SEG will be generated.

Understand the model in depth by reading our Evaluation section.

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 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 --fold 0. 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.

New: 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

--outdir <str>  # output directory
--fold <int> # use only a single model for inference 
--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

Python API

from mrsegmentator import inference
import os

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

inference.infer(images, outdir)

Change Path to Weights

MRSegmentator will automatically download its weights and save them in .conda/envs/<name>/lib/python3.11/site-packages/mrsegmentator/weights. This enables easy uninstallation including the weights, should you decide to clean your virtual environments.

If you have multiple environments set the MRSEG_WEIGHTS_PATH variable to prevent downloading multiple copies. Alternatively you can save the weights in a set location on your machine. For this you need to:

1. Download them from releases or move them from your conda environment
2. Unzip the files
3. Set the variable "MRSEG_WEIGHTS_PATH" to your weights directory (e.g.; export MRSEG_WEIGHTS_PATH="/home/user/weights)

How To Cite

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

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

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

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