WRDensity by Whiterabbit.ai

WRDensity is a software application that employs a deep learning algorithm to automatically analyze digital mammography images, providing an ACR BI-RADS breast density category to aid physicians in assessing breast tissue composition. The system, which processes DICOM images and outputs results to PACS/RIS, demonstrated strong performance with a quadratically-weighted Cohen's kappa of 0.90 when compared to a consensus of five expert radiologists for the four-class density task.

WRDensity by Whiterabbit.ai is a standalone software application designed to assist interpreting physicians in the assessment of breast tissue composition. It utilizes a deep learning algorithm to automatically analyze 'for presentation' data from compatible full-field digital mammography (FFDM) and digital breast tomosynthesis (DBT) systems. The software processes DICOM images transferred from a facility’s mammography system, PACS, or DICOM router. Its primary output is the Breast Density Level (BDL), a categorical assessment according to the American College of Radiology (ACR) BI-RADS Atlas 5th Edition breast density categories (A through D). A secondary output, Breast Density Level Probabilities (BDLP), provides the probabilities for each of the four density categories, offering insight into the device's confidence. These outputs are then sent to PACS and Radiology Information Systems (RIS) for review by radiologists as DICOM Secondary Capture Images or Structured Reports. The performance of WRDensity was evaluated through software verification and validation, including comparison against a consensus of five expert radiologists on a test dataset of 871 unique patient exams representing all compatible modalities. For the four-class BI-RADS task, WRDensity achieved a quadratically-weighted Cohen’s kappa of 0.90 (95% CI [0.88, 0.92]), indicating high agreement with the radiologist consensus. For a binary task (dense vs. non-dense tissue), it achieved a Cohen’s kappa of 0.88 (95% CI [0.85, 0.91]). The software also underwent consistency assessment, evaluating agreement between MLO and CC views of the same breast, and reproducibility testing. WRDensity is intended to provide adjunctive information and is not a diagnostic aid.

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 performance of WRDensity vary across different mammography machine vendors and models not explicitly listed as compatible, and what is the process for validating new systems?; What specific deep learning architecture is employed by WRDensity, and what were the key considerations for its selection and training in assessing breast density?; Beyond Cohen's kappa, what are the specific false positive and false negative rates for each individual BI-RADS category (A, B, C, D), and how might these impact clinical workflow or patient management decisions?; How does the 'Breast Density Level Probabilities (BDLP)' output, particularly the device's confidence level, correlate with instances of discordance between WRDensity and radiologist consensus, and can this be used to flag challenging cases?; What was the variability in breast density assessment among the five expert radiologists in the consensus panel, and how does WRDensity's performance compare to individual radiologist performance within that panel?