SAH Swin UNETR in the browser

The blood-segmentation network of the Río Hortega University Hospital group in Valladolid (García-García, Cepeda et al., World Neurosurg 2024), from their open SAH mortality-prediction pipeline: subarachnoid blood and its extensions, with volume, on non-contrast head CT. Everything runs on this computer.

Research and education use only

A research demonstration, not a medical device. This model is not validated or approved for clinical use here; do not use it for diagnosis, treatment or any other clinical decision. Segmentations can be wrong; always review them against the images.

Your images stay on this computer

Files are read and processed inside this browser tab on your GPU or CPU. No image data or results are uploaded. Only the page, the model files and the demo cases are downloaded, and the model is cached for later visits.

What you need

  • A recent desktop Chrome or Edge (Firefox and Safari may work on the CPU).
  • 8 GB RAM or more. WebGPU is fastest; on the CPU a scan takes one to several minutes.
  • A non-contrast axial head CT in Hounsfield units.
Compute
V

SAH Swin UNETR Río Hortega University Hospital, Valladolid · blood segmentation after aSAH

Blood on non-contrast head CT after aneurysmal subarachnoid hemorrhage (one class: subarachnoid blood with any intraventricular or intraparenchymal extension), with volume. The segmentation stage of the Valladolid SAH mortality-prediction pipeline.

García-García, Cepeda et al. · Valladolid & Augsburg
277 MB model download, cached after the first run. Runs on the CPU (WebAssembly) whatever the Compute setting: about 30 s per 128 × 128 × 64 window and usually two windows, so 1–2 minutes on a desktop; it needs about 2.3 GB of memory.
The network works in the authors’ 1 mm MNI CT template space. A scan from the RHUH dataset is already there and goes straight to the network; any other scan has its brain extracted (SAHVAI’s method; the original uses FSL BET) and is aligned to the template (rotation, shift and scale by mutual information; the original uses ANTs SyN), segmented, and mapped back so that volumes are measured on the scan itself.

Drop a head CT here

A zip of a DICOM study, a DICOM folder, loose DICOM files, a .nii / .nii.gz or a .nrrd volume. The RHUH demo cases are the dataset’s preprocessed, template-space scans; the BHSD cases are ordinary scans that are registered first.

reference: none · must share the scan's geometry; Dice is reported per class

Citation, licence and disclaimer

Model. García-García S, Cepeda S, Arrese I, Sarabia R. A fully automated pipeline using Swin transformers for deep learning-based blood segmentation on head computed tomography scans after aneurysmal subarachnoid hemorrhage. World Neurosurg 2024. doi:10.1016/j.wneu.2024.07.216 · García-García S, Cepeda S, Müller D, Mosteiro A, Torné R, Agudo S, de la Torre N, Arrese I, Sarabia R. Mortality prediction of patients with subarachnoid hemorrhage using a deep learning model based on an initial brain CT scan. Brain Sci 2024;14(1):10. doi:10.3390/brainsci14010010. Swin UNETR: Hatamizadeh A, et al. BrainLes 2021, LNCS 12962; MONAI (Apache-2.0).

Code and weights. github.com/smcch/Subarachnoid_Hemorrhage_segmentation_and_mortality_prediction (Dr Santiago Cepeda, Dr Dominik Müller and colleagues; © 2024 Hospital Río Hortega and University of Augsburg), checkpoint SWIN_UNETR_48_100_best_metric_model.pth linked from its README. Academic Non-Commercial Source Code License: academic, non-commercial research use only; “not for clinical decision-making”. Hosted here with the permission of Dr Santiago Cepeda (November 2026), with attribution to the publications above. The weights are converted to ONNX in full precision with no change (ONNX output matches PyTorch to 1.4 × 10⁻⁵, 0 voxels different). Two rewrites of the graph give the same arithmetic: the 2 × 2 × 2 transposed convolutions as 1 × 1 × 1 convolutions plus interleaving, and window attention computed in chunks of 64 windows so that it fits in the browser’s 4 GB memory. The 2 mm registration template is the repository’s 1 mm template, smoothed and subsampled. Its other model, the 3D DenseNet mortality predictor, is not distributed and not run here. CereBleed, the group’s newer multiclass tool (Cepeda S, et al. Neurosurgery 2026, doi:10.1227/neu.0000000000004020), is a separate model available at geibac.uva.es/cerebleed.

Demo data. RHUH cases 001 and 003 with their expert masks: Cepeda S, García-García S. Zenodo 2023, doi:10.5281/zenodo.8228847, CC BY 4.0. The model comes from the same hospital and these cases may have been used to develop it, so they show that the port is faithful, not how well the model generalises. BHSD cases as on the CLAIM page (labels: all blood).

How it runs here. As the repository’s inference_2.py (MONAI): Spacingd to 1.5 × 1.5 × 2 mm (bilinear) → z-score over the non-zero voxels → crop to the voxels above the mean → sliding window 128 × 128 × 64 with overlap 0.25 and constant blending → Swin UNETR (feature size 48) → logits back to the template (bilinear) → argmax. On RHUH cases 001–003 this port gives the same segmentation as the original Python pipeline (Dice 1.000; 63.0, 12.5 and 37.2 mL; Dice against the expert masks 0.77, 0.70 and 0.82). For other scans the original pipeline registers with ANTs SyN; in tests on three BHSD scans the original’s own result changed by Dice 0.79–0.93 between SyN and rigid registration, and this port agreed with its SyN result at Dice 0.77, 0.94 and 0.90. Volumes are measured on the scan; the original measures in template space, where they scale with head size.

Disclaimer. For research and education only. Not a medical device; not cleared or approved by the FDA, CE or any regulator; not validated for clinical use; must not be used for diagnosis, triage, treatment or any other clinical decision. Outputs can be wrong and must be reviewed against the images by a qualified reader. Provided "as is" without warranty of any kind. Images are processed locally in this browser and are not uploaded.

Log