3,500 non-contrast head CT scans with intracranial hemorrhage labels and hemorrhage-subtype segmentation

Overview

Our neuroradiology AI team is building a real-time triage algorithm for the emergency detection of intracranial hemorrhage (ICH) on non-contrast computed tomography (NCCT) of the brain. Acute blood appears hyperdense on NCCT (50–80 HU), making CT the first-line modality in stroke and trauma settings. We require axial NCCT series acquired at standard emergency-room protocols: tube voltage 120–140 kVp, slice thickness ≤5 mm (preferably 2.5 mm or thinner for posterior-fossa coverage), reconstructed with both brain window (WL 40 HU, WW 80 HU) and subdural window (WL 75 HU, WW 200 HU). Scout localizer images should be excluded from the de-identified package. Each scan must carry one or more of the following hemorrhage subtype labels: epidural hematoma (EDH), subdural hematoma (SDH), subarachnoid hemorrhage (SAH), intraparenchymal hemorrhage (IPH), or intraventricular hemorrhage (IVH). Pixel-level 2D or 3D segmentation masks indicating the hemorrhagic region are required for at least 60% of positive cases; the remainder may carry bounding-box annotations. Negative (no hemorrhage) cases should constitute 35–40% of the total dataset to reflect realistic emergency-room case mix and to enable balanced training. All DICOM files must be de-identified in compliance with GDPR Article 89 and HIPAA Safe Harbor, with UIDs re-mapped using a consistent pseudonymization scheme so longitudinal cases (admission plus follow-up) can be linked internally. Accompanying JSON sidecar files should encode subtype, hemorrhage volume estimate in milliliters, midline shift in millimeters, GCS score where available, and scan acquisition timestamp relative to symptom onset. NIfTI conversion is acceptable in addition to or instead of DICOM. Head CT volumes must undergo defacing or skull-stripping-based defacing prior to delivery to eliminate re-identification risk through facial reconstruction. Scanner diversity across 3T-equivalent protocols from GE Discovery, Siemens SOMATOM, and Philips Brilliance platforms is desirable. Volumetric DICOM series reconstructed at isotropic or near-isotropic resolution (≤2.5 mm) enable multiplanar reformatting for 3D lesion characterization. Annotation inter-rater reliability must achieve a minimum Dice similarity coefficient of 0.75 on hemorrhage masks across independent neuroradiologist reads, with adjudication by a third reader for discordant cases. QA exclusion criteria include scans with severe beam-hardening artifact from dental implants obscuring supratentorial structures, incomplete brain coverage, or imaging performed more than 24 hours after initial ictus without temporal metadata. The trained model will be deployed as a CE-marked and FDA 510(k)-pathway medical device for acute ICH flagging in radiology worklist prioritization systems. No patient data will leave the secure processing environment; a signed data-use agreement will be provided to every contributing institution.

Progress

Data Specifications

Use Cases

• Training and validating Medical imaging AI/ML models
• Benchmarking Medical imaging detection and segmentation algorithms
• Building de-identified Medical imaging research datasets for academic studies
• Augmenting existing Medical imaging datasets to reduce class imbalance