SIAM : Segment It All Model a head tissue segmentation model designed to be robust to contrast, resolution, and pathology. It can process any 3D human head volume (T1, T2, FLAIR, etc., and even CT).
The current version performs tissue segmentation, at 0.75 mm
-
Model 1 (-m 1) is performing a 39 regions segmentation task, as illustrated in the figure. It is an old version trained on 3 subjects, but may be interesting for extra-cerebral labels
-
Model 2 (-m 2) is described here, trained from 6 high-quality templates it shows great performance in healthy brain for tissue segmentation (12 brain tissue + head/skull/dura matter/vessel)
-
Model 3 (defautl or -m 0), extends siam's v2 robustness toward anatomical anomalies, which are segmented as an extra Label.
Valabregue, R., Khemir, I., Bardinet, E., Rousseau, F., Auzias, G. & Dorent R. (2026). SIAM : Head and Brain MRI Segmentation from Few High-Quality Templates via Synthetic Training. ArXiv. [https://arxiv.org/abs/2605.02737)
Note that you need to have a python3 installation for SIAM to work. pip version > 22 and setuptool > 61
-
Clone this repository:
git clone https://github.com/romainVala/SIAM
-
Go into the repository (the folder with the pyproject.toml file) and install: optionally create a conda env before the following commande
pip install -e .when testing on windows, I had to remove the -e flag ... (no idea why)
-
Model parameters will be downloaded to ~/siam_params/v0.x the first time you run an inference. If the installation is for multiple user, setup the environement variable
SIAM_MODEL_DIRand runpython /instal_dir/SIAMpred/download_model_weights.py. Then you only need to setupexport SIAM_MODEL_DIRbefore running the main commandsiam-pred
siam-pred -i INPUT_FILENAME INPUT_FILENAME must be a nifti file containing 3D volume data. 4D image sequences are not supported
siam-pred -i INPUT_FOLDER -o OUTPUT_PREFIX -m 3The above command will look for all nifti files in the INPUT_FOLDER
and save the predictions in a sub-folder containing the OUTPUT_PREFIX name.
if -o is not specify, result are store in the same folder, with a prefix
-m model number (default 3, int within [1 2 3] )
NO Interpolation: Use -voxel_size 0.75 for inputs with isotropic resolution in the range [0.6 1] mm. Adding this option
will avoid the standard process which consist of : 1) reslice input data to 0.75 mm (as it was the choosen training resolution)
2) interpolate back the predicted labels to the original input resolution. (Since we have to use nearest neighbor this will
induced a loss of smalll structures)
Removing interpolation is important for small vessel, but also for GM ... it worth the try ...
For very small baby brain, you need to scale the volume up, for the model to work.
You can achieve it, without resampling, by changing nifti header voxel size.
This will be done on the fly when using the -voxelsize x.x where x.x is a float for the
new (fake) voxel size. Typically for newborn brain we multiply the voxel size by 1.4, in order to get a
Total Intracranial Volume similar to an adult brain. The prediction result is then converted back to the
original resolution
Only use if you have near isotropic resolution
Use -nbthread 1 if you run with memory issues. This will reduce the number of thread for processing your input (default is 4)
To summarize all inputs parameters, refer to the help functionality:
siam-pred -hdocker pull romainvalabregue/siam Or with Singularity
singularity build siam_v0.3.simg docker://romainvalabregue/siamExample to run with gpu ()
singularity run --nv -B `pwd`:/data siam_v0.3.simg siam-pred -i /data/my_image.nii.gz For linux user, with limited disk space on $HOME or on /tmp, you can set following environement variable to location with enough space
export SINGULARITY_TMPDIR=/data/singularity/tpm
export SINGULARITY_CACHEDIR=/data/singularity/cache(may be also DOCKER_CONFIG= is you do not want to use the .docker in your home)
image will be ~10G and cache dir ~ 10G (can be removed after) tmp dir ~32G (but automaticaly deleted )
On Apple Silicon (M1–M4) the model can run on the integrated GPU via PyTorch's Metal backend. Either request it explicitly:
siam-pred -i INPUT_FILENAME -device mps -nbthread 1or just leave -device cuda as-is: when CUDA isn't available, SIAM now falls
back to MPS automatically (and only to CPU if MPS is also unavailable). A few
3D ops aren't implemented on Metal yet; PYTORCH_ENABLE_MPS_FALLBACK=1 is
set automatically so those transparently run on CPU.
Note that nnU-Net's sliding-window inference moves tiles between device and
host on non-CUDA backends, so MPS is faster than CPU but not as fast as CUDA.
On a 48 GB unified-memory machine, use -nbthread 1 or 2 to avoid OOM
kills from parallel worker processes.
Running with GPU requires more than 12 G on the GPU card
Execution time, and Memory usage will depend on the input image Field Of View (FOV). The input resolution do not matter too much since it is resliced to 0.75 mm resolution. So only the FOV will change the total datasize that will be feed into the network.
With a large FOV, covering the nec : 166x240x256 mm^3,
running with -device cpu -nbthread 8 took ~ 25 mn (seems to take ~ 20G of RAM)
running with -device cpu -nbthread 1 took ~ 2h20 (but sill ~17 G of RAM ... not sure why)
We thank the nnU-Net team for providing the training framework, and we built this inference tool based on HD-BET.
We thank B. Billot and E. Iglesias for their original SynthSeg method: training on synthetic data enables robust, contrast-agnostic segmentation models. It worth the try, and we are still working to improve it, adding more tissue, being robust to anomalies
Thanks to Fernando and its torchio lib ! my favorite 3D augmentation tools which we used to build the generative model.
Thanks goes to these wonderful people (emoji key):
 Benoît Béranger 💻 |
 Chris Rorden 💻 |
This project follows the all-contributors specification. Contributions of any kind welcome!