LST-AI - Deep Learning Ensemble for Accurate MS Lesion Segmentation

2026-05-06: LST-AI has received the NeuroImage: Clinical Editors’ Choice Award 2025 🏆

Welcome to our codebase for LST-AI, the deep learning-based successor of the original Lesion Segmentation Toolbox (LST) by Schmidt et al. LST-AI was collaboratively developed by the Department of Neurology and Department of Neuroradiology, Klinikum rechts der Isar at the Technical University of Munich, and the Department of Computer Science at the Technical University of Munich.

Disclaimer: LST-AI is a research-only tool for MS Lesion Segmentation and has not been validated, licensed or approved for any clinical usage.

What is different, why or when should I switch?!

• LST-AI is an advanced deep learning-based extension of the original LST with improved performance and additional features.
• LST-AI constitutes a completely new framework and has been developed from scratch.
• While LST depends on MATLAB, we offer LST-AI as a python-based tool which makes it available to the whole community.
• We suggest using LST or LST-AI according to your type of data:
  • A 3D T1-weighted and a 3D FLAIR sequence are available: (new) LST-AI
  • Only a 3D FLAIR sequence is available: (old) LST (LPA)
  • Only a 3D T1-weighted sequence is available: not covered by any LST(-AI) version
  • If a 3D T1-weighted and a non-3D FLAIR sequence are available, please try both (new) LST-AI or (old) LST (LGA or LPA)

Installation

LST-AI installs from PyPI. Inference runs in PyTorch — there is no TensorFlow, no ONNX Runtime and, since v2.0.0, no onnx dependency either: the released weights ship as .pt checkpoints. greedy (registration), HD-BET (skull stripping) and FastSurfer's dependencies install as wheels on every supported platform, so nothing is compiled.

python3 -m venv lst_env && source lst_env/bin/activate
pip install lst-ai

While 2.0.0 is in its release-candidate phase, pip needs to be told that a pre-release is acceptable: pip install --pre lst-ai (or pin it explicitly, pip install lst-ai==2.0.0rc1). Once the final 2.0.0 is published, the plain command above is all there is.

The model bundle and atlas are downloaded automatically on first run; they land next to the installed lst_ai package when that directory is writable (any venv install), and fall back to ~/.cache/lst_ai when it is not. Set LST_AI_DATA_DIR to override the location. Lesion annotation runs on FastSurfer, which the install brings along — see Lesion annotation with FastSurfer. Nothing has to be set up by hand.

pip install --index-url https://download.pytorch.org/whl/cu126 torch

uv venv --python 3.12 lst_env && source lst_env/bin/activate
uv pip install lst-ai

python3 -m venv lst_env && source lst_env/bin/activate
git clone https://github.com/CompImg/LST-AI/ && cd LST-AI
pip install -e .

What happened to TensorFlow?

Inference was reimplemented natively in PyTorch for v2.0.0. The released weights are unchanged — every tensor matches the released .h5 bit-for-bit; nothing was retrained. Validated against the released TensorFlow models on 10 subjects, full ensemble: Dice agreement 0.99565, lesion volume difference −0.12 %, lesion Dice vs consensus ground truth 0.6829 (TF) vs 0.6835 (PyTorch). For scale: running the identical code in a container rather than a virtualenv moves the result more (Dice 0.883) than the framework swap does, because greedy's registration is nondeterministic run to run.

Details, the retired ONNX backend's drift, and guidance for comparing against earlier results: docs/pytorch-reimplementation.md. The release validation checklist, including what remains untested: docs/testing.md — further "worked for me" reports, especially from aarch64 and CPU-only machines, are genuinely useful.

Lesion annotation with FastSurfer

Lesions are assigned to anatomical regions using a FastSurfer segmentation of the T1. Every mode except --segment_only runs it, so FastSurfer is a hard requirement of a normal lst run — and both the pip install and the Docker image bring their own pinned copy with them, found without any FASTSURFER_HOME or PATH setup. A FastSurfer already on the machine is deliberately ignored, so an annotation never depends on what a given machine happens to have lying around.

How the install works, pre-fetching checkpoints for offline machines, and LST_AI_SKIP_FASTSURFER: docs/fastsurfer.md.

Usage

Once installed, LST-AI is a command line tool. It expects zipped NIFTIs (*.nii.gz) as input and assumes the input images are not skull-stripped (if yours are, pass --stripped, or segmentation performance will be severely affected).

lst --t1 t1.nii.gz --flair flair.nii.gz --output /mnt/data/lst/results --temp /mnt/data/lst/processing

--t1, --flair and --output are required. --temp keeps the intermediate files (e.g. skull-stripped images in MNI152 space); without it a temporary directory is used and removed afterwards.

Modes

1. Default — Segmentation + Annotation: provide the T1w and FLAIR images; LST-AI segments and annotates lesions according to McDonald criteria.
2. Segmentation only (--segment_only): only the binary segmentation mask is saved.
3. Annotation only (--annotate_only): annotate an existing binary mask, provided in FLAIR space via --existing_seg /path/to/binary/mask.

Common settings

• --device: integer GPU ID, or cpu. The default is 0, i.e. GPU 0. On a machine without a GPU — including the CPU Docker image — you must pass --device cpu.
• --stripped: inputs are already skull-stripped (both of them — a mixture is not supported).
• --threshold: threshold applied to the ensemble's lesion probability map to produce the binary mask (default 0.5).

Dockerfile and Dockerhub

For offline or reproducible use, LST-AI ships as Docker images with every model weight baked in — a container runs with no network access. Images are published to Docker Hub under jqmcginnis/lst-ai, one CUDA and one CPU flavour per release, each as a multi-arch manifest (linux/amd64 + linux/arm64):

docker pull jqmcginnis/lst-ai:v2.0.0        # CUDA flavour
docker pull jqmcginnis/lst-ai:v2.0.0-cpu    # CPU flavour

latest and latest-cpu track the newest stable release. Pre-releases are published only under their explicit tag (e.g. v2.0.0rc1, v2.0.0rc1-cpu) and never move latest, so a plain docker pull jqmcginnis/lst-ai cannot land on a candidate. The legacy jqmcginnis/lst-ai_cpu repository and the v1.x tags remain for the old TensorFlow-based images.

To build the images yourself — including behind a TLS-inspecting hospital or university proxy — see docker/Readme.md.

Running the container

Bind-mount your input and output directories (absolute paths — Docker requires them) and pass the usual lst arguments:

# GPU image
docker run --gpus all \
  -v /home/ginnis/lst_in:/in -v /home/ginnis/lst_out:/out -v /home/ginnis/lst_temp:/temp \
  lst-ai:gpu --t1 /in/t1.nii.gz --flair /in/flair3d.nii.gz --output /out --temp /temp

# CPU image: no --gpus, and --device cpu is required
docker run \
  -v /home/ginnis/lst_in:/in -v /home/ginnis/lst_out:/out -v /home/ginnis/lst_temp:/temp \
  lst-ai:cpu --t1 /in/t1.nii.gz --flair /in/flair3d.nii.gz --output /out --temp /temp --device cpu

Results appear directly in the bind-mounted output directory.

docker run -u $(id -u):$(id -g) --gpus all -v /home/ginnis/lst_in:/in -v /home/ginnis/lst_out:/out lst-ai:gpu --t1 /in/t1.nii.gz --flair /in/flair3d.nii.gz --output /out

Extending and modifying LST-AI

We invite you to tailor LST-AI to your pipeline and application, please have a look at our sources.

BIDS Compliance with LST-AI

To ensure maximum flexibility for our user base, LST-AI does not natively enforce BIDS-compliant file-naming conventions. This decision allows users to work seamlessly with both BIDS and non-BIDS datasets.

However, for those who wish to utilize LST-AI within a BIDS-compliant workflow, we have provided an example repository that demonstrates the integration of LST-AI with BIDS-compliant data. This example reflects the BIDS-compliant usage of LST-AI that we are currently using in our internal database.

Citation

Please consider citing LST-AI to support the development:

@article{wiltgen2024lst,
  title={LST-AI: A deep learning ensemble for accurate MS lesion segmentation},
  author={Wiltgen, Tun and McGinnis, Julian and Schlaeger, Sarah and Kofler, Florian and Voon, CuiCi and Berthele, Achim and Bischl, Daria and Grundl, Lioba and Will, Nikolaus and Metz, Marie and others},
  journal={NeuroImage: Clinical},
  pages={103611},
  year={2024},
  publisher={Elsevier}
}

Further, please also credit greedy and HD-BET, used for preprocessing the image data, and — unless you run --segment_only — FastSurfer, which provides the anatomical parcellation lesions are assigned to.

@article{yushkevich2016ic,
  title={IC-P-174: Fast Automatic Segmentation of Hippocampal Subfields and Medial Temporal Lobe Subregions In 3 Tesla and 7 Tesla T2-Weighted MRI},
  author={Yushkevich, Paul A and Pluta, John and Wang, Hongzhi and Wisse, Laura EM and Das, Sandhitsu and Wolk, David},
  journal={Alzheimer's \& Dementia},
  volume={12},
  pages={P126--P127},
  year={2016},
  publisher={Wiley Online Library}
}

@article{isensee2019automated,
  title={Automated brain extraction of multisequence MRI using artificial neural networks},
  author={Isensee, Fabian and Schell, Marianne and Pflueger, Irada and Brugnara, Gianluca and Bonekamp, David and Neuberger, Ulf and Wick, Antje and Schlemmer, Heinz-Peter and Heiland, Sabine and Wick, Wolfgang and others},
  journal={Human brain mapping},
  volume={40},
  number={17},
  pages={4952--4964},
  year={2019},
  publisher={Wiley Online Library}
}

@article{henschel2020fastsurfer,
  title={FastSurfer -- A fast and accurate deep learning based neuroimaging pipeline},
  author={Henschel, Leonie and Conjeti, Sailesh and Estrada, Santiago and Diers, Kersten and Fischl, Bruce and Reuter, Martin},
  journal={NeuroImage},
  volume={219},
  pages={117012},
  year={2020},
  publisher={Elsevier},
  doi={10.1016/j.neuroimage.2020.117012}
}

@article{henschel2022fastsurfervinn,
  title={FastSurferVINN: Building resolution-independence into deep learning segmentation methods -- A solution for HighRes brain MRI},
  author={Henschel, Leonie and K{\"u}gler, David and Reuter, Martin},
  journal={NeuroImage},
  volume={251},
  pages={118933},
  year={2022},
  publisher={Elsevier},
  doi={10.1016/j.neuroimage.2022.118933}
}