DeepCT

DeepCT is an artificial intelligence-driven software designed to expedite the identification of suspected intracranial hemorrhage in non-contrast head CT scans by providing rapid notifications to specialists. Utilizing a deep residual convolutional neural network (ResNet) trained on extensive radiological data, the system analyzes DICOM images to improve triage efficiency in acute care settings without replacing definitive diagnostic procedures or full patient evaluation.

DeepCT is a software-only medical device designed to assist clinicians in hospital emergency rooms by rapidly identifying potential cases of intracranial hemorrhage (ICH) from non-contrast CT brain scans. It functions as a notification-only tool, operating in parallel with standard care workflows to flag images for specialist review. The system comprises two main components: an Image Forwarding Software that transmits DICOM images and displays notifications, and an Image Processing and Analysis Server that houses the core artificial intelligence algorithm. This server receives, processes, and analyzes the CT images, sending results back to the forwarding software, and is intended for use by trained clinicians within hospital networks to identify and communicate images of specific patients to a specialist, independent of the standard care workflow. The core technology of DeepCT is a 34-layer deep residual convolutional neural network (ResNet), trained using the PyTorch library on a large dataset of over 21,000 retrospective head CT scans collected between 2007 and 2017. This AI model is specifically engineered to detect image characteristics consistent with ICH by analyzing non-contrast CT images of the brain acquired in acute settings. Performance validation was conducted through a retrospective, multicenter, multinational study involving 260 cases from five clinical sites, demonstrating a sensitivity of 93.8% and a specificity of 92.3% in identifying ICH, both exceeding the 80% performance goal. The average image processing and notification time was approximately 30.6 seconds, highlighting its capability for rapid triage support, though it is not intended for diagnostic use beyond notification and should not replace full patient evaluation.

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. How does the DeepCT algorithm differentiate between different types or severities of intracranial hemorrhage (e.g., subdural, epidural, subarachnoid, intraparenchymal)?; What were the specific inclusion/exclusion criteria for the 21,603 patient CT scans used in the training dataset, and how might biases in this dataset affect real-world performance across diverse populations?; Beyond sensitivity and specificity, what is the positive predictive value (PPV) and negative predictive value (NPV) of DeepCT in a typical emergency department setting with varying ICH prevalence rates?; How does the system handle images with significant artifacts or poor image quality, and what is the failure rate or an "unable to analyze" output under such conditions?; What specific IT infrastructure and security protocols are required for the seamless and secure integration of DeepCT's Image Forwarding Software and Image Processing and Analysis Server within a hospital's existing network and PACS?