MammoScreen

MammoScreen is a radiological computer-assisted detection software that employs deep learning algorithms trained on biopsy-proven examples to aid physicians in identifying suspicious calcifications and soft tissue lesions in screening FFDM. The system processes mammograms to highlight findings and assign a 1-10 suspicion score, with clinical studies demonstrating improved radiologist detection performance when using the software.

MammoScreen is a software-only medical device designed as a concurrent reading aid for physicians interpreting screening Full-Field Digital Mammograms (FFDM). Its primary purpose is to assist in identifying focal findings suspicious for breast cancer, such as calcifications and soft tissue lesions, and to assess their level of suspicion. The system comprises a processing server and a web interface, utilizing advanced image analysis algorithms, including deep learning machine learning components. These algorithms are trained on extensive databases of biopsy-proven breast cancer examples, benign lesions, and normal tissue. MammoScreen processes FFDM images and provides output as CAD marks highlighting suspicious findings, along with a "MammoScreen Score" on a 1-10 scale indicating increasing suspicion, with findings scoring 5 or higher initially prioritized for review. Standalone performance testing demonstrated MammoScreen's capability in the absence of a clinician, achieving a combined ROC AUC of 0.883 at the mammogram level and 0.911 at the breast level across images from Hologic and GE devices. The system showed notable performance for detecting both soft tissue lesions (LROC AUC 0.877) and calcifications (LROC AUC 0.942) at the finding level. A retrospective multi-reader multi-case clinical study involving 14 radiologists and 240 mammograms found that using MammoScreen as an aid improved radiologists' cancer detection performance, increasing the average Area Under the ROC Curve (AUC) from 0.77 (unaided) to 0.80 (aided). This improvement was observed for 11 out of 14 radiologists and was statistically significant, though average reading time per case slightly increased with MammoScreen assistance. The system's standalone performance (AUC 0.79) was found to be non-inferior to the average performance of unaided radiologists.

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 MammoScreen algorithm's performance (sensitivity/specificity) vary across different breast density categories, and was this factored into the training dataset?; The clinical study showed an average increase in reading time; what is the distribution of this increase among radiologists, and are there specific workflow integrations that could mitigate this?; What were the characteristics of the false positives generated by MammoScreen in the standalone testing, and how do they compare to typical false positives identified by radiologists alone?; Given the performance differences observed between Hologic and GE devices in standalone testing (e.g., LROC AUC for soft tissue lesions was 0.837 for Hologic vs. 0.900 for GE), what steps are taken to ensure consistent performance across a wider range of mammography systems?; The MammoScreen Score is a 1-10 scale. How was the calibration of this score to the actual probability of malignancy established and validated, particularly for the thresholds (e.g., MS5) used for default display?