Neuro Lesion Inpainting Tool (neuroLIT) 🔥

Overview

This repository contains sourcecode and documentation related to our publication FastSurfer-LIT: Lesion Inpainting Tool for Whole Brain MRI Segmentation With Tumors, Cavities and Abnormalities. This tool can inpaint lesions independent of their shape or appearance for further downstream analysis. This allows subsequent analysis to run as if no lesion were present, enabling for example, FastSurfer's whole brain segmentation in cases with large brain tumors or surgical cavities. The extended documentation is located at DeepMI.org/neurolit.

Quickstart

Docker / singularity

git clone https://github.com/Deep-MI/neurolit
cd neurolit
./neurolit/scripts/run_lit_containerized.sh --input_image T1w.nii.gz --lesion_mask lesion_mask.nii.gz --output_directory output_directory
# Add --singularity to use singularity instead of docker

Pip

pip install neurolit
lit-inpainting --input_image T1w.nii.gz --lesion_mask lesion_mask.nii.gz --output_directory output_directory

How to run neuroLIT

We recommend using containerization in combination with the neurolit/scripts/run_lit_containerized.sh wrapper script. This will automatically pull the Docker image from Docker Hub or create a Singularity image cache and run the neuroLIT inpainting. Additional lit-inpainting flags such as --dilate 2 are forwarded through the wrapper to the container entrypoint. If you do not pass --tag, the wrapper selects the newest numeric release tag available on Docker Hub (for example 0.6.0), rather than Docker's floating latest tag. We also have a pip release of neuroLIT.

Running neuroLIT

The most straight forward way of doing the inpainting is just providing

1. The T1w image
2. The lesion mask
3. An output directory
4. (optional) The number times to dilate the lesion mask (default: 0)

./neurolit/scripts/run_lit_containerized.sh --input_image T1w.nii.gz --mask_image lesion_mask.nii.gz --output_directory output_directory --dilate 2

The default is to use Docker. Add the --singularity flag to use Singularity or Apptainer instead:

./neurolit/scripts/run_lit_containerized.sh --singularity --input_image T1w.nii.gz --mask_image lesion_mask.nii.gz --output_directory output_directory

On first use, the image is pulled from Docker Hub — no root required. The image is cached in containerization/deepmi_lit_<version>.sif and reused on subsequent runs.

If you already have a .sif file, pass it directly to skip the pull:

./neurolit/scripts/run_lit_containerized.sh --singularity_image /path/to/deepmi_lit.sif --input_image T1w.nii.gz --mask_image lesion_mask.nii.gz --output_directory output_directory

Both singularity and apptainer (the HPC-standard open-source fork) are supported automatically.

The outputs will be placed in the output directory in the folder inpainting_volumes and contain

• The inpainted T1w image (inpainting_result.nii.gz)
• The (dilated) mask used for inpainting in the same space as the input image (inpainting_mask.nii.gz)
• The conformed original image (inpainting_original_image.nii.gz)

We recommend performing dilation, since undersegmentation can negatively impact the performance of the inpainting, while oversegmentation should not have significant impact.

If the source image was isotropic, the output images should have the same resolution as the input image and the area outside of the lesion mask should be preserved, except for a robust rescaling of the intensity values.

Installation from PyPI

The same interface as above can be accessed from pypi:

# Install the package
pip install neurolit

# Download model checkpoints (recommended - download once, ~700MB)
lit-download-models

# Run neuroLIT
lit-inpainting --input_image T1w.nii.gz --lesion_mask lesion_mask.nii.gz --output_directory output_directory --dilate 2

Note: If you skip the lit-download-models step, models will be automatically downloaded on first use.

Expected Runtimes

Runtimes below are for a typical whole-brain T1w volume (1 mm isotropic), measured with neuroLIT v0.6.0:

Hardware	Runtime
NVIDIA GeForce RTX 5070 Ti (16 GB)	~17 minutes
Intel Xeon w5-2465X (CPU only)	~5.5 hours

GPU is strongly recommended for interactive use. CPU is feasible for batch/overnight processing.

Integration with FreeSurfer/FastSurfer

neuroLIT is being integrated into FastSurfer for whole brain segmentation and cortical surface reconstruction of images with lesions.

For standalone usage with FreeSurfer/FastSurfer, neuroLIT provides post-processing scripts for integrating lesions into FastSurfer/FreeSurfer outputs. We recommend using the unified lit-postprocessing script which handles mapping the lesion mask to multiple segmentation files, running volume statistics (segstats), and performing surface masking.

Postprocessing

This script modifies FastSurfer/FreeSurfer segmentations and surface annotation files, and the corresponading statistics (.stats) files.

# Setup paths
export FASTSURFER_HOME=/path/to/FastSurfer
export FREESURFER_HOME=/path/to/freesurfer
source $FREESURFER_HOME/SetUpFreeSurfer.sh

# Run unified postprocessing
lit-postprocessing \
    -sid SUBJECT_ID \
    -sd /path/to/subjects_dir

Overview:

• Requirements: Automatically checks for FastSurfer/FreeSurfer installations using $FASTSURFER_HOME and $FREESURFER_HOME. FreeSurfer is required for surface postprocessing.
• Dynamic Configuration: Uses segstats_config.json for volumetric stats and surfstats_config.json for surface stats. By default, all relevant FastSurfer annotation and segmentation files are modified.
• Adjacent Label Reports: Generates adjacency reports for lesion segmentations, allowing users to assess which regions are affected or replaced by the lesion.

The current postprocessing workflow updates the primary FastSurfer outputs in place, keeps the pre-lesion versions as .lit backups or mapped backup files, and writes lesion-specific reports such as stats/aparc.DKTatlas+aseg.lesion_report.txt, stats/aseg.lesion_report.txt, and stats/lesion_impact_summary.yaml.

Training

The training script can be found here. The same docker image can be used for training, but you need to mount the training data directory using the -v flag. Note that training data are expected to be conformed (with the script conform.py).

Documentation

Comprehensive documentation is available in the doc/ directory and will be made public at DeepMI.org/neurolit.

Documentation includes:

• Installation guides
• Usage tutorials
• API reference
• Contributing guidelines

References

If you use neuroLIT for research publications, please cite:

Pollak C, Kuegler D, Bauer T, Rueber T, Reuter M, FastSurfer-LIT: Lesion Inpainting Tool for Whole Brain MRI Segmentation with Tumors, Cavities and Abnormalities, Imaging Neuroscience 2025. https://doi.org/10.1162/imag_a_00446