Saige-Density

Saige-Density is a medical image processing software that employs deep learning algorithms to analyze digital mammograms (FFDM and DBT) and provide an automated ACR BI-RADS breast density assessment. It is designed to serve as an adjunctive tool for radiologists, aiming to aid in the consistent and objective evaluation of breast tissue composition, with its performance validated through a retrospective multi-site study achieving 81.28% four-class accuracy against expert consensus.

Saige-Density is a Software as a Medical Device designed to assist interpreting physicians in assessing breast tissue composition from digital mammograms. It utilizes deep learning techniques to analyze full-field digital mammography (FFDM) and digital breast tomosynthesis (DBT) images for female patients aged 35 and older. The software processes these images and generates a single study-level breast density category (A, B, C, or D) according to the ACR BI-RADS Atlas 5th Edition guidelines. This output, provided as DICOM structured report and secondary capture objects, is intended as adjunctive information and explicitly not as a standalone diagnostic tool, meaning clinical decisions should not rely solely on its analysis. The scientific basis of Saige-Density lies in its proprietary deep learning algorithms, which were trained on diverse datasets from various US geographic locations, including racially diverse populations. To validate its performance, a multi-site retrospective study was conducted using 796 mammogram cases (6,170 images) from five US breast imaging centers, distinct from those used for algorithm development. The software's accuracy in assigning a four-class breast density category (A, B, C, or D) was evaluated against a consensus ground truth established by five expert radiologists. This study demonstrated an accuracy of 81.28% (95% CI: 78.42, 83.84) for the four-class categorization, with further details provided for two-class categorization.

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-Density's performance vary across different mammography machine manufacturers (e.g., Hologic, GE) and imaging modalities (FFDM vs. DBT) beyond the aggregate results presented?; What specific deep learning architecture is employed by Saige-Density, and how was the training data preprocessed and augmented to achieve the reported performance, especially considering racial diversity in datasets?; How does the inter-observer variability among the five expert radiologists who established the ground truth compare to the variability between Saige-Density and the consensus ground truth?; What is the performance of Saige-Density on challenging cases, such as images with artifacts, implants, or prior surgeries, and were such cases explicitly included or excluded in the validation study?; Can the system provide a confidence score or probability associated with its assigned BI-RADS density category to give clinicians a better understanding of the AI's certainty?