CuraRad-ICH

CuraRad-ICH is a software tool that employs a deep learning-based 3D classification framework to analyze adult non-contrast head CT scans for suspected acute intracranial hemorrhage. Its purpose is to enhance radiologist workflow by providing a rapid, automated case-level prediction to aid in the prioritization and triage of critical ICH cases for expedited clinical assessment.

CuraRad-ICH is a software workflow tool designed to assist in prioritizing the clinical assessment of adult non-contrast head CT scans suspected of acute intracranial hemorrhage (ICH). It functions as a Software as a Medical Device (SaMD), utilizing deep learning algorithms to analyze these CT images by identifying potential ICH based on the higher density (Hounsfield Units) of blood compared to brain tissue; the software provides a case-level binary prediction (suspected ICH: yes/no) to a PACS/workstation, enabling worklist prioritization or triage for radiologists, with its core technology involving an end-to-end trainable 3D classification framework for automated ICH detection. Performance was evaluated in a retrospective, blinded, multisite clinical validation study on 388 CT studies, where the software demonstrated an ICH detection sensitivity of 90.6% and a specificity of 93.1%. Negative Predictive Value (NPV) remained high across various prevalence rates (e.g., 99.9% at 1% prevalence), while Positive Predictive Value (PPV) varied with prevalence (e.g., 70% at 15% prevalence), and the average system processing time per study was 43 seconds with no significant performance differences based on CT slice thickness, detector rows, or scanner manufacturers; importantly, CuraRad-ICH is not intended for stand-alone diagnostic decision-making, to direct attention to specific image portions, or to rule out hemorrhage, and must not replace full clinical assessment.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. What specific deep learning architecture and training dataset characteristics (e.g., size, diversity of ICH types, scanner vendors) underpin CuraRad-ICH's 3D classification framework?; How does CuraRad-ICH's performance vary with subtle or atypical presentations of intracranial hemorrhage, and are there specific ICH subtypes where its sensitivity or specificity is notably different?; Beyond binary ICH presence, does the algorithm provide any localization information or quantitative assessment (e.g., estimated volume) that could further aid in triage, even if not for primary diagnosis?; What are the real-world data integration challenges and typical setup times when integrating CuraRad-ICH with existing hospital Picture Archiving and Communication Systems (PACS) and worklist applications?; Given the high Negative Predictive Value (NPV), what are the clinical implications and potential workflow changes if a 'negative' flag from CuraRad-ICH is used to de-prioritize scans, and what safeguards are recommended?