Lunit INSIGHT CXR Triage

Lunit INSIGHT CXR Triage is an AI-powered software designed to analyze adult chest X-rays for suspected pleural effusion and pneumothorax, flagging critical cases within PACS/workstations to assist medical specialists in prioritizing their review workflow. The system utilizes a vendor-agnostic deep learning algorithm on DICOM images to detect these time-sensitive conditions, with its performance in identifying features suggestive of these findings validated through nonclinical and clinical studies demonstrating high ROC AUC, sensitivity, and specificity.

Lunit INSIGHT CXR Triage is a radiological computer-assisted software tool designed for triage and notification in adult chest X-ray analysis. It employs artificial intelligence algorithms to identify pre-specified critical findings, specifically pleural effusion and/or pneumothorax, on frontal chest X-ray images. The software integrates with existing Picture Archiving and Communication Systems (PACS) or workstations, flagging potentially critical images to enable prioritized review by appropriately trained medical specialists qualified to interpret chest radiographs; it does not alter the order of cases or remove them from the reading queue. It operates as a passive notification system, meaning it does not send proactive alerts, but rather groups suspicious exams for potential earlier review within the standard care workflow, and is explicitly not intended for standalone clinical decision-making or directing attention to specific image portions beyond indicating the presence of the target findings. The device processes de-identified DICOM chest radiographs, utilizing a vendor-agnostic artificial intelligence algorithm to analyze each image for features suggestive of pleural effusion or pneumothorax, providing a case-level output ('flag' or blank) to the PACS/workstation. It can identify multiple critical findings per radiograph, and users can selectively enable notifications for each condition. When deployed with imaging equipment, analysis occurs automatically post-acquisition; an on-device notification also informs technologists (15 minutes post-interpretation by the specialist) which cases were prioritized by the AI, without providing diagnostic data. Nonclinical validation on 1,385 images yielded ROC AUCs of 0.9864 for pleural effusion (94.29% sensitivity, 95.72% specificity) and 0.9973 for pneumothorax (96.08% sensitivity, 99.14% specificity). Clinical studies using 1,708 images from NIH and Indian datasets demonstrated ROC AUCs of 0.9686 for pleural effusion (89.86% sensitivity, 93.48% specificity) and 0.9630 for pneumothorax (88.92% sensitivity, 90.51% specificity), with an average analysis and notification time of approximately 20-21 seconds per finding.

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 the severity or subtlety of pleural effusion and pneumothorax (e.g., small vs. large findings)?; What were the specific demographic characteristics (age, sex, comorbidities) and image quality parameters of the NIH and India datasets used in the clinical validation, and how might these affect generalizability?; How does the system handle images with multiple concurrent critical and non-critical pathologies, and does the presence of other findings impact the accuracy for pleural effusion or pneumothorax detection?; What is the defined threshold for the 'flag' notification, and can users adjust this threshold based on their clinical setting's tolerance for false positives versus false negatives?; Considering the 15-minute delay for technologist notification *after* specialist interpretation, what is the intended actionable outcome or utility of this specific notification for the technologist?