MammoScreen® (3)

MammoScreen 3 is an artificial intelligence-based software intended to assist physicians in interpreting screening mammograms by detecting suspicious findings, providing a numerical suspicion score, and characterizing lesion type and location. Utilizing deep learning algorithms trained on extensive datasets, including prior mammograms, the device aims to enhance diagnostic accuracy and streamline the reporting process in breast cancer screening.

MammoScreen 3 is an artificial intelligence-powered software device designed as a concurrent reading and reporting aid for physicians interpreting screening mammograms, including Full-Field Digital Mammography (FFDM) and Digital Breast Tomosynthesis (DBT) images, with the option to incorporate prior examinations in its analysis. The system utilizes deep learning algorithms to process mammograms, detect findings suspicious for breast cancer, and assign a MammoScreen ScoreTM from one to ten, indicating the level of suspicion. It outputs graphical marks of findings, characterizes lesions as mass/asymmetry, distortion, or calcifications, and provides detailed localization information (quadrant, depth, distance from nipple). This information can be integrated into DICOM viewers and reporting systems, but patient management decisions should not solely rely on MammoScreen 3's analysis. A multi-reader multi-case (MRMC) clinical study demonstrated that radiologists using MammoScreen 3 significantly improved their diagnostic performance in breast cancer detection, with average Area Under the Curve (AUC) increasing from 0.797 to 0.871, sensitivity from 0.706 to 0.793, and specificity from 0.815 to 0.836. As a standalone system, MammoScreen 3 showed an AUC of 0.883 in the MRMC study, proving superior to unaided radiologists and non-inferior to radiologists aided by the device. A larger standalone performance study on 7,544 exams yielded an overall AUROC of 0.927, sensitivity of 0.912, and specificity of 0.727, with robust performance across various subgroups including breast density, lesion type, and patient age. Compared to its predicate, MammoScreen 3 offers updated algorithmic components for improved detection accuracy, more detailed lesion characterization, and enhanced reporting aid features.

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 MammoScreen 3's performance vary specifically with different Digital Breast Tomosynthesis (DBT) acquisition protocols or reconstruction algorithms from various equipment vendors, given the standalone performance data primarily mentioned Hologic systems?; What is the quantified incremental benefit of using prior examinations in the analysis by MammoScreen 3, and how does the time interval between current and prior exams impact the AI's detection and characterization accuracy for different lesion types and breast densities?; The multi-reader multi-case study indicated an overall improvement in radiologist performance; what was the extent of inter-reader variability in this performance improvement, and did specific radiologist characteristics (e.g., years of experience, sub-specialization focus) correlate with greater or lesser benefit from using MammoScreen 3?; How does the MammoScreen ScoreTM (ranging from 1 to 10) precisely correlate with established BI-RADS assessment categories, and what is the recommended clinical workflow for managing discordant findings where the AI's suspicion score significantly differs from the radiologist's initial assessment?; Considering the detailed standalone performance for lesion type assessment (e.g., Positive Percentage Agreement for distortion was 0.544), what specific algorithmic enhancements or data augmentation strategies are being explored to improve the characterization accuracy for more subtle or challenging findings like architectural distortions?