Research and education use only
Not a medical device and not validated for clinical use. The signs are found by rules written from their published visual definitions, applied to an AI segmentation: treat every finding as a candidate for a reader to confirm, and an absent sign as not excluded. Do not use it for diagnosis or treatment decisions.
Your images stay on this device
Files are read and processed inside this browser tab. No image data or results are uploaded. Only the page, the segmentation model (26 MB, cached), the Yale expansion model (33 MB, when it runs) and the demo cases are downloaded.
What you need
- An axial head CT with a spontaneous ICH: non-contrast for the NCCT signs, CT angiography for spot-sign candidates. DICOM zip or folder, NIfTI or NRRD.
- A recent desktop Chrome or Edge; WebGPU is fastest. About 30 s to 3 min per scan.
Hematoma expansion signs NCCT signs · CTA spot sign · BAT, BRAIN and target expansion scores · DeepMedX patches
Step 1 segments the hemorrhage with SAHVAI (BrainHemoAI Hybrid 2D/3D UNet) on the 0.5 × 0.5 × 5 mm grid that the DeepMedX pipeline also uses. Step 2 reads the expansion signs inside and around the largest intraparenchymal hematoma. Step 3 combines them with the clinical details you enter.
Clinical details for the scores (optional, stay on this device)
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. From a study the axial non-contrast head series nearest 5 mm is used; if the only series is a CTA, it is used and spot-sign candidates are looked for instead of the NCCT signs.
The signs click a card to jump to its slice · set Reader to confirm or reject a finding; scores use the reader's call when given
Scores
Expansion prediction model Yale (Tran et al., npj Digit Med 2024) · runs after the signs
Expansion on a follow-up CT conventional HE and revised HE (rHE, with IVH growth)
Drop the follow-up CT of the same patient (or type the volumes) to classify expansion. Conventional HE: ICH growth > 6 mL or > 33% (Dowlatshahi et al., Neurology 2011). Revised HE (rHE): ICH growth ≥ 6 mL or ≥ 33%, or IVH growth ≥ 1 mL, or new IVH (Yogendrakumar et al., Stroke 2020;51:1120–1127; used by Ning, Yu et al., Sci Rep 2025). Usual follow-up: 24 h (up to 72 h).
Drop the follow-up CT here
DeepMedX end-to-end pipeline Yu et al., npj Digit Med 2026
Method, citations and disclaimer
Segmentation. SAHVAI Web (BrainHemoAI Hybrid 2D/3D UNet; Hu P, et al. NeuroImage 2023;279:120321, doi:10.1016/j.neuroimage.2023.120321), labels SAH, IPH and IVH on a 0.5 × 0.5 mm, 5 mm grid. The hematoma is the largest 26-connected IPH component (optionally with IVH).
Sign definitions. Morotti A, Boulouis G, Dowlatshahi D, et al. Standards for detecting, interpreting, and reporting noncontrast computed tomographic markers of intracerebral hemorrhage expansion. Ann Neurol 2019;86:480–492, and the original reports cited on each card. Review: Liu S, Fu W, Tang X, Fan Y. CT-based prediction of hematoma expansion and adverse outcomes after intracerebral hemorrhage. GeroScience 2026, doi:10.1007/s11357-026-02478-3.
Scores. BAT: Morotti A, et al. Stroke 2018;49:1163–1169. BRAIN: Wang X, et al. Stroke 2015;46:376–381 (PMID 25503550). Target expansion score: Morotti A, et al. Neurology 2026;107(9):e218675 (PMID 42879724). Spot sign score: Delgado Almandoz JE, et al. Stroke 2009;40:2994–3000.
Yale expansion model. Tran AT, Zeevi T, Haider SP, et al., Payabvash S. Uncertainty-aware deep-learning model for prediction of supratentorial hematoma expansion from admission non-contrast head computed tomography scan. npj Digit Med 2024;7:26, doi:10.1038/s41746-024-01007-w; code and weights github.com/anhtrnyaleedu/HE (ATACH-2 + Yale, n = 634; HE ≥ 6 mL AUC 0.81 in high-confidence test cases). Converted to ONNX without weight changes; preprocessing re-implemented from model512v9.ipynb. © 2024 Yale University, BSD 3-Clause License; hosted with the authors’ permission. Inclusion here does not imply endorsement by Yale University or the authors. See also Tran et al., AJNR 2025;46:1404–1411 (robustness).
Expansion definitions. Dowlatshahi D, et al. Neurology 2011;76:1238–1244 (conventional); Yogendrakumar V, et al. Redefining hematoma expansion with the inclusion of intraventricular hemorrhage growth. Stroke 2020;51:1120–1127 (revised); Ning Y, Yu Q, et al. Sci Rep 2025;15:32021.
DeepMedX pipeline. Yu Q, Fan X, Li J, et al. An end-to-end deep learning pipeline for hematoma expansion prediction in spontaneous intracerebral hemorrhage based on non-contrast computed tomography. npj Digit Med 2026;9:39, doi:10.1038/s41746-025-02213-w; code github.com/DeepMedX/sICH_HE (non-commercial research licence). Their region-of-interest step (crop2DROI.ipynb) is re-implemented here in JavaScript; their U-Mamba, Diffusion-UKAN and ViT weights are not public, so their classifier does not run here.
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. The detectors have been checked on synthetic phantoms and inspected on a few public cases, not validated against expert sign reads. Do not use for diagnosis, prognosis or treatment decisions.
CT signs of hematoma expansion reference
The figure of Liu et al. (GeroScience 2026, Fig. 2) shows ten panels and nine signs (the blend sign twice): three contrast-extravasation signs (spot, iodine, leakage) and six NCCT signs (blend, black hole, island, satellite, hypodensity, swirl). Irregular shape, heterogeneous density and fluid level complete the NCCT set of the 2019 consensus (Morotti et al., Ann Neurol). Schematics drawn for this page (anterior at the top).
How the signs are read
- Hematoma. The largest connected IPH component of the SAHVAI segmentation; each axial section is hole-filled so encapsulated low-density areas stay inside it. Density signs are read on sections with at least 30% of the largest section's area (end sections are partial-volumed) and not on sections with dense metal (a coil or clip, ≥ 2000 HU) in the brain; a catheter near the clot is flagged.
- Low-density regions. The CT is lightly smoothed (σ ≈ 0.5 mm). Pixels ≥ 12 HU below the section's hyperdense reference (75th percentile of the clot core) and ≥ 12 HU are candidates; a region must reach ≥ 1.5 mm inside the clot edge, be ≥ 7 mm² and ≥ 2 mm wide. It is encapsulated if growing it 1.5 mm through low-density pixels does not reach the clot edge. Its HU difference is measured against the hyperdense ring 0.5–2 mm around it.
- Density signs. Hypodensity: encapsulated, ≥ 15 HU difference. Black hole: encapsulated, ≥ 28 HU. Swirl: encapsulated, iso- or hypodense to brain (≤ brain + 8 HU), on ≥ 2 contiguous slices. Blend: open to the edge, ≥ 18 HU, ≥ 15 mm², ≤ 60% of the section, a ≥ 5 mm interface with a sharp edge (≥ 4 HU/mm), and still denser than brain. Fluid level: the low-density part lies wholly above a straight (RMS ≤ 0.75 mm), near-horizontal (≤ 10°) interface ≥ 8 mm long spanning half the clot. Heterogeneous density: ≥ 3 low-density foci (≥ 2 mm²) on the largest section.
- Shape signs. Lobules: what an opening with a disk (radius 0.35 × the section's equivalent radius, 2–8 mm) removes from the section; bubble- or sprout-like ones are ≤ 80 mm² with a neck narrower than themselves. Separate small hematomas: other blood labels or compact (elongation ≤ 3) foci of 50–100 HU, ≥ 2 mm², ≥ 4 mm inside the brain, away from metal and calcification. Satellite: a separate focus < 10 mm across, 1–20 mm from the clot. Island: ≥ 3 separate, or ≥ 4 separate plus connected, on one section. Irregular shape: ≥ 2 lobules or separate foci within 10 mm on the largest section.
- Contrast. A scan with ≥ 2 mL of 130–600 HU voxels inside the brain is treated as contrast-enhanced: NCCT signs are not read, and foci ≥ 120 HU, ≥ 1.5 mm, inside the clot and not touching its edge are spot-sign candidates. Thin CTA slices are averaged into 5 mm slabs, which lowers small spots' HU.
Yale model. Their pipeline as published: brain window 0–80 HU, skull stripped, padded or cropped to 512 × 512 × 48 (slices resampled to 5 mm when there are more than 48), SegResNet by 256 × 256 × 32 sliding window; then mask and CT resampled to 1 mm, the slices with blood ± 16 mm, the CT within 40 mm of the clot as a second channel, resized and centre-cropped to 128 × 128 × 96, 3D DenseNet121 with Monte-Carlo dropout. Checked against the authors' code in Python: same probability (0.175 vs 0.174) and mask volume (80.0 mL) on the BHSD demo.
Self-test: node tools/hesigns_selftest.mjs injects each sign into a synthetic hematoma and checks it is found (and that a plain hematoma has none). On public demo cases the outputs were inspected by eye; no expert-read sign data set was available.