QOCA image Smart CXR Image Processing System

The QOCA® image Smart CXR Image Processing System is an AI-powered software designed to analyze adult chest X-rays for suspected pneumothorax, aiming to improve radiology workflow efficiency by prioritizing cases for review. Its deep learning algorithm, validated on diverse datasets, demonstrates high sensitivity and specificity in identifying potential pneumothorax, integrating with hospital PACS to provide timely notifications to radiologists.

The QOCA® image Smart CXR Image Processing System is a software as a medical device (SaMD) designed to assist radiologists by prioritizing the review of adult chest X-ray images suspected of containing pneumothorax. It utilizes a locked artificial intelligence algorithm based on deep learning technology to analyze DICOM images automatically pushed from a hospital's Picture Archiving and Communication System (PACS). The system is web-based, supports multiple users, and provides features for case sorting and image viewing. Its methodology involves analyzing posteroanterior (PA) or anteroposterior (AP) erect view chest X-rays to identify potential pneumothorax cases, subsequently making a notation next to these cases in the worklist to alert radiologists. The system is intended as an adjunctive tool and does not provide a diagnosis or indicate specific anomalies on the image, serving solely to flag cases for prioritized review. The system's performance was validated using independent training and testing datasets. The training dataset incorporated diverse patient demographics and imaging parameters, including data from the US National Institutes of Health (NIH) to enhance US population representativeness. Performance assessment was conducted through two pivotal studies using data from the MIMIC dataset (for US population generalizability) and a Taiwanese hospital dataset (for imaging equipment generalizability), both truthed by three radiologists. Across these assessment datasets, the system achieved an overall Area Under the Curve (AUC) of 97.8%, with a sensitivity of 92.5% and specificity of 94.0%. Specifically, on the MIMIC dataset (3,105 radiographs), it demonstrated an AUC of 97.7%, sensitivity of 93.7%, and specificity of 93.3%. On the Taiwanese dataset (2,947 radiographs), it achieved an AUC of 97.4%, sensitivity of 91.7%, and specificity of 94.9%. The average time for the device to analyze a study and send a notification was 4.94 seconds.

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 algorithm's performance vary across different subtypes or severities of pneumothorax (e.g., small vs. tension pneumothorax)?; What were the specific characteristics of the NIH dataset incorporated into the training data, and how did this inclusion demonstrably improve US patient population representativeness?; Given the system flags cases at a case-level without indicating specific anomalies, what is the typical workflow impact and time-saving benefit for radiologists in a real-world clinical setting beyond the notification time?; Were there any notable differences in performance when analyzing PA view versus AP erect view images, or across different X-ray imaging equipment beyond the two models listed in the Taiwanese dataset?; How does the "locked" nature of the AI algorithm accommodate potential future retraining or updates as new data or imaging technologies become available, and what is the process for such updates?