ClearRead Xray Pneumothorax

ClearRead Xray Pneumothorax is an artificial intelligence-driven software designed to aid radiologists by rapidly triaging adult frontal chest X-rays, prioritizing those with suspected pneumothoraces of 5mm or larger for expedited review. The system utilizes machine learning and image processing on DICOM images, and its clinical effectiveness for notification and prioritization has been demonstrated through performance metrics like high AUC, sensitivity, and specificity on independent datasets.

ClearRead Xray Pneumothorax is a computer-assisted triage software designed for trained professionals, such as radiologists, to prioritize adult frontal chest X-rays suspected of containing a pneumothorax 5mm or larger. The system analyzes digital DICOM images, requiring both lungs to be in the field of view, and uses an artificial intelligence algorithm based on machine learning and image processing to identify features suggestive of pneumothorax. Its output is delivered via a DICOM network interface to a PACS or workstation, facilitating worklist prioritization by highlighting potentially urgent cases. This tool functions as an adjunctive notification system and is not intended for direct diagnosis, replacement of full patient evaluation, or use with plain film X-rays. The device's clinical efficacy was validated using an independent dataset comprising over 1000 retrospective adult chest X-rays, sourced from the MIMIC-CXR and Georgetown University Medical Center databases, including more than 400 true positive and 600 true negative pneumothorax cases. Performance was assessed by the Area Under the ROC Curve (AUC), sensitivity, specificity, and time-to-notification, with ground truth established by a panel of three board-certified radiologists. The clinical study demonstrated an AUC of 0.974, sensitivity of 0.922, specificity of 0.951, and an average notification time of 9.73 seconds, indicating its effectiveness in prioritizing images for review. Algorithm development utilized both real and synthetically generated pneumothorax data, with internal testing showing comparable performance (AUC 0.975).

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 with pneumothorax size, particularly for those at or near the 5mm detection threshold specified in the indications for use?; What specific image preprocessing and normalization techniques are employed by the 'normalization component' to mitigate variability across different X-ray manufacturers and patient characteristics encountered in clinical practice?; Given the use of synthetic data in algorithm training, what measures were taken to ensure the synthetic pneumothoraces accurately represent the diverse morphological presentations and subtlety of real-world pneumothoraces?; How does the presence of common comorbidities, such as pleural effusion or lung opacity as detailed in Table 4, impact the algorithm's detection accuracy (sensitivity and specificity) for pneumothorax?; What were the specific criteria for selecting the 0.5 probability threshold for the internal test set's operating point, and how does varying this threshold affect the trade-off between sensitivity and specificity for triage purposes in a clinical workflow?