Saige-Dx

Saige-Dx is an artificial intelligence software that analyzes digital breast tomosynthesis and associated 2D mammography images to detect suspicious lesions and calcifications, providing finding-level and case-level suspicion scores to support radiologists. Performance data from reader and standalone studies demonstrate its ability to enhance cancer detection accuracy, notably by increasing radiologists' AUC and sensitivity in interpreting screening mammograms.

Saige-Dx is an artificial intelligence-driven software tool designed to assist interpreting physicians in the analysis of screening digital breast tomosynthesis (DBT) mammograms for women aged 35 and older. The system processes DBT image stacks along with accompanying 2D images (full-field digital mammography and/or synthetic 2D images) to identify the presence or absence of soft tissue lesions and calcifications potentially indicative of cancer. It takes DICOM files as input and generates outputs including bounding boxes around detected findings, a 'Finding Suspicion Level' for each, and an overall 'Case Suspicion Level'. These results are presented in DICOM Structured Report format for image overlay and as a DICOM Secondary Capture object containing a summary report, functioning as a concurrent reading aid. The device's efficacy was evaluated through two retrospective studies using data distinct from its AI training sets, which were sourced from diverse US and UK locations. A multi-reader multi-center study involving 18 MQSA qualified radiologists demonstrated that using Saige-Dx significantly improved their breast cancer detection capabilities. The average reader Area Under the Curve (AUC) increased from 0.865 without Saige-Dx to 0.925 with its aid, accompanied by an average sensitivity increase of 8.8% and a specificity increase of 0.9%. These improvements were consistent across different breast densities, patient ages, racial/ethnic groups, lesion types (soft tissue densities and calcifications), and radiologist specializations. A separate standalone performance study involving 1304 cases showed Saige-Dx achieved an AUC of 0.930, underscoring its strong performance relative to unaided readings.

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 Saige-Dx's performance vary when used with digital breast tomosynthesis hardware from manufacturers other than Hologic, which was specified in the reader study?; What are the specific architectural features or training data characteristics of the Saige-Dx AI algorithm that enable it to effectively identify both recalled and non-recalled (initially missed) cancers?; Can the 'Finding Suspicion Level' and 'Case Suspicion Level' be quantitatively correlated with pathological outcomes, and how robust is this correlation across diverse patient demographics and breast densities not fully detailed in the summary?; Beyond AUC and sensitivity/specificity, what was the measured impact of Saige-Dx on radiologists' reading times, particularly comparing breast imaging specialists versus general radiologists?; Given the reported increase in specificity by 0.9% in the reader study, what is the projected impact on false positive rates and patient recall rates in a large-scale, real-world screening population?