Research use only
A research demonstration, not a medical device. Not validated or approved for clinical use; do not use it for diagnosis or treatment. Segmentations and 3D surfaces can be wrong; check them against the slices.
Your images stay on this computer
Scans are read and segmented inside this browser tab on your GPU or CPU. Nothing is uploaded. The model (about 26 MB) is downloaded once and cached.
Serial scans (4D)
- Drop several zips, folders or volumes at once, or add scans one at a time.
- DICOM studies are separated and dated from their headers; NIfTI / NRRD files take a date from the file name or from the table.
- Each scan takes about 30 s on WebGPU, a few minutes on the CPU.
Drop one head CT for a 3D map — or 2–6 serial CTs for SAHVAI-4D
Zips of DICOM studies, DICOM folders, loose DICOM files, .nii / .nii.gz or .nrrd volumes. From each study the axial non-contrast head series nearest 5 mm is used.
Drag to rotate · scroll to zoom · right-drag to pan
Layers
SAHVAI-4D
Steps through the scans in date order with the same camera. Scans are centred on the brain, not registered to each other.
Slices
Coronal and sagittal views are reformatted from 5 mm axial slices, so they look stepped.
Blood volume over time
SAH here is all subarachnoid blood found by the model; the ABC/2 SAHV of the eSAH score covers 5 cisternal spaces only. Volumes from different scanners or slice thicknesses are not strictly comparable.
Log
How it works
- Segmentation. Each scan runs through the same pipeline as SAHVAI Web: resampling to 0.5 × 0.5 × 5 mm, brain extraction and the Hybrid 2D/3D UNet of Hu et al. (BrainHemoAI weights, ONNX Runtime Web), giving SAH, IPH and IVH labels and volumes.
- Skull removal. The 3D brain is the intracranial soft tissue the pipeline already finds (−20 to 100 HU, enclosed by the skull), so bone, scalp and the scanner table never appear.
- Surfaces. Each label is smoothed lightly, the 5 mm slices are interpolated to about 1.25 mm and a marching-cubes surface is drawn with three.js. The ventricles are an approximate, HU-based context layer (CSF density deep in the brain plus IVH), not a model output. Volumes always come from the voxel labels, never from the surfaces.
- 4D. Studies are separated by StudyInstanceUID and ordered by acquisition date. Volumes are plotted against date; the SAH clearance rate and a log-linear half-life are shown when there are enough dated scans.