Saige-Q

Saige-Q is an artificial intelligence software tool designed to aid radiologists in prioritizing full-field digital mammography (FFDM) and digital breast tomosynthesis (DBT) screening exams for review. It employs deep neural networks to analyze DICOM images and generate a code indicating the likelihood of suspicious findings, thereby facilitating worklist triage without offering diagnostic information.

Saige-Q is a medical imaging software designed to function as a workflow tool for radiologists by prioritizing screening mammograms. It utilizes artificial intelligence, specifically deep neural networks, to analyze Digital Breast Tomosynthesis (DBT) and Full-Field Digital Mammography (FFDM) studies. The software processes DICOM files from a single screening study, checks for appropriateness, and then the AI algorithm generates a 'Saige-Q code'. This code indicates the software's suspicion of findings suggestive of breast cancer. This output can be integrated into PACS, EPR, or RIS worklists, allowing radiologists to reorder their review queue, potentially addressing suspicious cases sooner. Saige-Q is intended for passive notification only and does not provide diagnostic information, nor is it meant to replace a radiologist's full review or be used for stand-alone clinical decision-making; it is compatible with Hologic mammography equipment. Performance data for Saige-Q was established through two retrospective, blinded, multi-center studies for FFDM and DBT mammograms. In the FFDM study (501 malignant, 832 normal exams), Saige-Q achieved an AUC of 0.966, with a specificity of 92.2% at 86.9% sensitivity. For the DBT study (517 malignant, 1011 normal exams), it achieved an AUC of 0.985, with a specificity of 98.3% at 86.9% sensitivity. These results met the primary endpoints, which aimed for sensitivity and specificity above BCSC data benchmarks. Sub-analyses showed consistent performance across lesion types, breast densities, age, and lesion sizes. Median processing times were 15.5 seconds for FFDM and 196.8 seconds for DBT, deemed clinically acceptable.

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-Q's performance specifically compare on FFDM versus DBT images for cases with subtle or early-stage malignancies?; What are the characteristics of the datasets used to train and validate the deep neural networks, particularly regarding diversity in patient populations and imaging equipment beyond Hologic?; How does the 'Suspicious' code threshold for Saige-Q align with different radiologists' individual sensitivity/specificity preferences or institutional protocols for recall rates?; What is the typical processing time for a DBT study in a real-world clinical environment with varying network loads and server capacities, and how might this impact workflow for high-volume centers?; Are there any plans to integrate Saige-Q with features that might provide localization information for the suspicious findings, even if not for primary diagnostic use?