https://atlas.rsna.org/schemas/2025-11/model.json

Detection of Acute or Subacute Hemorrhage on Noncontrast CT Head Scans

https://dx.doi.org/10.1148/ryai.210115

1.0

Eli Gibson; email: moc.sreenihtlaeh-snemeis@nosbig.ile

Supported/financially supported by Siemens Healthineers.

Anonymized NCCT head volumes were retrospectively collected from 10 centers with approval from their respective ethics committees, which waived the requirement for written informed consent, in compliance with HIPAA.

Deep convolutional neural network comprising dual-orientation (axial and coronal) feature-extractor image-to-image subnetworks, segmentation subnetworks, and orientation-specific classifiers with a common classifier head; trained with voxelwise segmentation loss and study-level sigmoid binary cross-entropy for ICH and five subtypes; additional uncertainty modeling via Dempster-Shafer subjective logic estimating Beta distribution parameters.

Prototype developed and evaluated in a retrospective study; not stated as publicly available software.

Automatic detection, subtyping, and segmentation of acute or subacute intracranial hemorrhage on NCCT; provides confidence scores that identify high-confidence subsets with higher accuracy and enables improved simulated radiology worklist prioritization (reduced average report turnaround time).

Patients’ triage; clinical decision support systems; workflow optimization (worklist prioritization).

Operating points selected to maximize Youden index on held-out development data; classification scores calibrated so the operating-point threshold maps to 0.5 for entropy-based confidence scoring.

Fully automated image analysis producing study-level classifications, subtype labels, segmentations, volume estimates, and confidence scores to support clinician decision making.

Evaluation of adult patients (>18 years) undergoing noncontrast head CT to detect, subtype, and locate acute or subacute intracranial hemorrhage; intended for use in radiology workflow environments.

Axial and coronal noncontrast head CT volumes (NCCT).

Preprocess by automatic alignment to a standard reference frame using anatomic landmarks; resample to 1-mm in-plane and 4-mm out-of-plane resolution; normalize CT intensities by mapping −45 to 155 HU to the range 0–1. If no coronal reconstruction is present, resample axial volume to coronal grid.

Generalization gap observed for epidural hemorrhage (low prevalence in development data, ~0.7%); confidence scores depend partly on hemorrhage volume; anonymization limited assessment of factors influencing performance; urgency depends on broader clinical context not modeled; confidence scores derived from statistical models may not match radiologist confidence; RTAT improvements demonstrated via a simplified queuing model and require prospective clinical validation; calibration required across datasets from different centers.

Use confidence scores to triage into high-confidence positive, uncertain, and high-confidence negative categories to enhance worklist prioritization beyond binary classification.