Research and education use only
A research demonstration, not a medical device and not a triage system. It is not validated or approved for clinical use; do not use it for diagnosis, triage, transfer or treatment decisions. It can miss hemorrhage and can flag normal scans; a qualified reader must review every scan.
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 model files and the demo cases are downloaded, and the model is cached for later visits.
What you need
- A non-contrast axial head CT: a DICOM zip or folder, loose DICOM files, NIfTI or NRRD.
- A recent Chrome, Edge or Safari on a computer, tablet or phone. The default model is a 14 MB download and needs about 1 GB of memory.
HEADS-UP detection and 5-subtype classification · RSNA 2019
Step 1 of the SAHVAI Web suite: is there intracranial hemorrhage, what type, and on which slices? Step 2: send the scan to the segmentation tool that suits the type found, for volumes and localisation.
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. On a phone, use Choose file(s) with a zip or NIfTI. The axial non-contrast head series is picked from a study; slices thinner than 4 mm are averaged to 5 mm.
Citation, licence and disclaimer
HEADS-UP. Salman S, Gu Q, Dherin B, Reddy S, Vanderboom P, Sharma R, Lancaster L, Tawk R, Freeman WD. Hemorrhage Evaluation and Detector System for Underserved Populations: HEADS-UP. Mayo Clin Proc Digit Health 2023;1(4):547–556. doi:10.1016/j.mcpdig.2023.08.009. Developed with the Google Advanced Solutions Lab (Dr Benoit Dherin, Sanjana Reddy). The 2023 HEADS-UP model was a cloud AutoML classifier; this browser version replaces it with open networks so that it can run on the device.
Networks. The 2D CNNs of the 1st-place solution of the RSNA 2019 Intracranial Hemorrhage Detection challenge (team of Tao Shen, Xiyue Wang and colleagues): Wang X, Shen T, Yang S, Lan J, Xu Y, Wang M, Zhang J, Han X. A deep learning algorithm for automatic detection and classification of acute intracranial hemorrhages in head CT scans. NeuroImage Clin 2021;32:102785. doi:10.1016/j.nicl.2021.102785. Code and weights: github.com/SeuTao/RSNA2019_Intracranial-Hemorrhage-Detection (fold-0 checkpoints of DenseNet121 512 px, DenseNet169 256 px and SE-ResNeXt101 256 px), converted to ONNX with the weights unchanged; stored as 16-bit numbers to halve the download (probabilities change by less than 0.0001). The solution's sequence models, which refine the slice predictions, have no published weights and are not included. The converted weights are hosted with the written permission of Tao Shen (1 November 2026), with attribution as above; see LICENSE.md.
Training data. Flanders AE, Prevedello LM, Shih G, et al. Construction of a machine learning dataset through collaboration: the RSNA 2019 Brain CT Hemorrhage Challenge. Radiol Artif Intell 2020;2(3):e190211. doi:10.1148/ryai.2020190211. The RSNA Brain CT Hemorrhage Dataset (Copyright RSNA, 2019) is licensed for non-commercial use only; so is this tool.
How it runs here. As the solution's prepare_data.py and predict.py: each slice in the brain window (the DICOM's own, else 40/80 HU), min-max scaled to 8 bits, resized to 512 px, stacked with the slices above and below as three channels, centre-cropped to 80 % and resized back, then through the network(s). The study result for each type is its highest slice probability. In this port the network input matches OpenCV's (Python) to within one grey level in about 1 pixel in 10,000, and the slice probabilities match the Python pipeline to within 0.00001 (BHSD demo case). Localisation on the image is approximate: hyperdense voxels (45–100 HU) inside each detected type's class-activation hot spot. It is not a segmentation; use the segmentation tools for volumes. The 80 % centre crop means the outer edge of the field of view is not seen by the networks.
Demo data. BHSD case SAH_002 (Wu B, et al. MLMI 2023, as distributed with the CLAIM demo), MISTIE/CLEAR case 01 (Hanley D, et al., Johns Hopkins, doi:10.7281/T1/CZDPSX, CC BY-NC 4.0), BLAST-CT scan_0 (Apache-2.0) and the BrainHemoAI head.nrrd (MIT), as on the other SAHVAI Web pages.
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.