openjdata/siamize
siamize - native C++/ONNX port of SIAM v0.3 brain segmentation (CUDA, OpenCL, TensorRT, CoreML)
119
openjdata/siamize repository overview
siamize — native C++ port of SIAM v0.3 brain/head MRI segmentation
- Copyright: (C) Qianqian Fang (2026) <q.fang at neu.edu>
- License: Apache License, Version 2.0
- Version: 0.2.0
- GitHub: https://github.com/NeuroJSON/siamize
- Upstream: https://github.com/romainVala/SIAM — SIAM v0.3 by Valabregue, Khemir, Bardinet, Rousseau, Auzias & Dorent (2026), arXiv:2605.02737
- Acknowledgement: This project uses resources and data formats developed as part of the NeuroJSON project supported by US National Institute of Health (NIH) grant U24-NS124027.
Native C++ port of SIAM v0.3 (Segment It All Model) for head/brain tissue segmentation — runs with no PyTorch / nnU-Net at inference time. One image, two GPU backends:
| Binary (entrypoint) | Backend | Hardware |
|---|---|---|
siamize (default) | ONNX Runtime + CUDA EP | NVIDIA GPU, or CPU |
siamize-opencl | MNN + OpenCL | NVIDIA / AMD / Intel GPU, or CPU |
Input: NIfTI (.nii/.nii.gz) or JNIfTI (.jnii/.bnii). Output: a uint8
labelmap (18 SIAM classes) or a 4D tissue-probability map. Fold weights
auto-download from NeuroJSON on first run — mount a volume at /cache to
keep them.
Quick start
--gpus all exposes the GPU (and the NVIDIA OpenCL driver). Mount your data at
/data and a named cache volume at /cache.
# NVIDIA GPU, ONNX Runtime / CUDA — 5-fold ensemble (default entrypoint)
docker run --rm --gpus all -v "$PWD":/data -v siamize-cache:/cache \
openjdata/siamize:v2026.6 \
-i /data/in.nii.gz -o /data/labels.nii.gz -M 0,1,2,3,4 -c cuda
# Single fold (faster, lighter)
docker run --rm --gpus all -v "$PWD":/data -v siamize-cache:/cache \
openjdata/siamize:v2026.6 \
-i /data/in.nii.gz -o /data/labels.nii.gz -M 0 -c cuda
# Vendor-neutral GPU via MNN / OpenCL (NVIDIA / AMD / Intel) — override entrypoint
docker run --rm --gpus all -v "$PWD":/data -v siamize-cache:/cache \
--entrypoint siamize-opencl openjdata/siamize:v2026.6 \
-i /data/in.nii.gz -o /data/labels.nii.gz -M 0 -c opencl
# CPU only (no GPU needed — drop --gpus)
docker run --rm -v "$PWD":/data -v siamize-cache:/cache \
openjdata/siamize:v2026.6 \
-i /data/in.nii.gz -o /data/labels.nii.gz -M 0 -c cpu
# List the OpenCL devices the container sees (then pick one with -G N)
docker run --rm --gpus all --entrypoint siamize-opencl \
openjdata/siamize:v2026.6 --list-gpu
# 4D tissue-probability map (float32 softmax) instead of a labelmap
docker run --rm --gpus all -v "$PWD":/data -v siamize-cache:/cache \
openjdata/siamize:v2026.6 \
-i /data/in.nii.gz -o /data/tpm.nii.gz -M 0,1,2,3,4 -c cuda --tpm
# SPM12-style 6-class output (GM, WM, CSF, Bone, Soft, Air)
docker run --rm --gpus all -v "$PWD":/data -v siamize-cache:/cache \
openjdata/siamize:v2026.6 \
-i /data/in.nii.gz -o /data/spm.nii.gz -M 0 -c cuda -C spm
# JNIfTI output (JSON / binary, for the NeuroJSON ecosystem)
docker run --rm --gpus all -v "$PWD":/data -v siamize-cache:/cache \
openjdata/siamize:v2026.6 \
-i /data/in.nii.gz -o /data/labels.jnii -M 0 -c cuda -F jnii
Common options
| Flag | Meaning |
|---|---|
-i FILE | input volume (.nii[.gz] / .jnii / .bnii) |
-o FILE | output file (extension picks the format, or use -F) |
-M 0,1,2,3,4 | fold weights (digits 0–4 = the 5-fold ensemble; auto-downloaded) |
-c {auto|cuda|cpu} (ORT) / {auto|opencl|cpu} (siamize-opencl) | backend; auto picks GPU then falls back to CPU |
-G N | GPU device. ORT: CUDA index. siamize-opencl: 1-based device from --list-gpu |
--list-gpu | (siamize-opencl) list OpenCL devices and exit |
-t N | CPU threads (default auto) |
-P ZxYxX | sliding-window patch (default 256x256x192; smaller = less memory) |
-u S | target isotropic spacing in mm (default 0.75) |
-C N|spm | output classes: 18 (default) or spm (6 TPM channels) |
--tpm | write a 4D float32 tissue-probability map instead of a labelmap |
-F nii|jnii|bnii | output container |
--lowmem | force the low-memory preset (smaller patch + tighter knobs) |
Run --help (default entrypoint) for the full list:
docker run --rm openjdata/siamize:v2026.6 --help
Volumes & environment
| Mount / var | Purpose |
|---|---|
-v "$PWD":/data | your input/output files |
-v siamize-cache:/cache | persists auto-downloaded fold weights (+ OpenCL tuning cache) across runs |
--gpus all | required for any GPU run — exposes the GPU and the NVIDIA OpenCL driver |
To confirm an OpenCL run is on the GPU (not a silent CPU fallback): the log
shows device=opencl with no OpenCL init error, and nvidia-smi shows
GPU activity. CPU-only hosts work with -c cpu and no --gpus.
Tags
Calendar-versioned vYYYY.M (e.g. v2026.6). The image is CUDA 12 + cuDNN 9
based; the MNN/OpenCL binary is static and vendor-neutral.
Links
- Source / docs: https://github.com/NeuroJSON/siamize
- Upstream SIAM v0.3: https://github.com/romainVala/SIAM — arXiv:2605.02737
- Please cite the SIAM paper if you use this in your work.
Acknowledgement: uses resources and data formats from the NeuroJSON project, supported by US NIH grant U24-NS124027.
Tag summary
Content type
Image
Digest
sha256:6b044b2dc…
Size
2.2 GB
Last updated
4 months ago
docker pull openjdata/siamize:v2026.6