DeepTek CXR Analyzer v1.0

DeepTek CXR Analyzer v1.0 is a software device utilizing deep learning algorithms to analyze adult and transitional adolescent chest radiographs, identifying and highlighting suspicious regions of interest in the Lungs, Pleura, Cardiac, and Hardware categories as a concurrent reading aid for radiologists. Its clinical utility is supported by standalone performance assessments and a multi-reader, multi-case study demonstrating a statistically significant improvement in radiologists' ability to detect and localize abnormalities when using the software.

The DeepTek CXR Analyzer v1.0 is a computer-assisted detection (CADe) software device designed to assist radiologists by analyzing adult and transitional adolescent (18 to <22 years) frontal chest radiographs. It employs deep learning algorithms to identify, categorize, and highlight suspicious Regions of Interest (ROIs) within four predefined categories: Lungs, Pleura, Cardiac, and Hardware. The system integrates with Picture Archiving and Communication Systems (PACS), receiving studies in DICOM format, processing them, and returning a secondary capture DICOM image. This output DICOM includes the original radiograph overlaid with color-coded bounding boxes for detected suspicious ROIs, or a message indicating no suspicious ROIs were found. The device is intended as a concurrent reading aid and does not provide a diagnosis or treatment recommendations; radiologists are instructed to review original images alongside the CADe output. The software was trained on a dataset of over 100,000 chest X-ray images from 30 sites in India, across more than 15 modality manufacturers. Standalone performance was evaluated using a dataset of 3,000 scans (from NIH Chest X-ray Database and Segmed Insight Platform), showing aggregate AUROC of 0.974 for detection and aggregate wAFROC-FOM of 0.920 for localization. A clinical performance assessment, a multiple reader, multiple case (MRMC) retrospective study involving 24 U.S. board-certified radiologists and 300 chest radiographs, demonstrated that reader performance significantly improved with the aid of the device. The aggregate wAFROC-FOM for aided reads was 0.893, compared to 0.821 for unaided reads (p<0.0001), with all 24 readers showing improved wAFROC-FOM for localizing suspicious ROIs across all categories when aided. Sensitivity and specificity also improved with device assistance.

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 device's performance vary across different patient demographics (e.g., specific age subgroups within the 18+ range, detailed ethnicity breakdowns) beyond the summary provided for the Segmed Insight Platform dataset?; What are the specific types of pathological findings or hardware characteristics within the broad categories of 'Lungs', 'Pleura', 'Cardiac', and 'Hardware' for which the algorithm demonstrates peak versus lowest performance?; Considering the training dataset originated from India and the standalone/clinical testing utilized U.S. datasets, what specific steps were taken during model development and validation to ensure robustness against variations in imaging equipment, patient populations, and disease prevalence between these regions?; What is the computational time required for the DeepTek CXR Analyzer to process an individual chest radiograph and return the annotated DICOM file to PACS, and how does this integrate into time-sensitive clinical workflows?; Were there any notable differences in performance improvement (aided vs. unaided wAFROC-FOM) among radiologists with varying levels of experience in the MRMC study, and if so, what were they?