2917. Artificial Intelligence Platform for Determining Breast Density in Digital Breast Tomosynthesis Mammography: How Does it Perform?
Authors * Denotes Presenting Author
  1. Liane Philpotts *; Yale School of Medicine
  2. Peter Szabla; Visage Imaging, Inc
  3. Mingde Lin; Yale School of Medicine
Artificial Intelligence (AI) tools are being rapidly developed to aid in radiology daily workflow and interpretation. Data from actual clinical use of these tools is important to assess how systems perform and how radiologists interact with AI. Breast density assessment has been a subjective decision for which there is potential interobserver variability. The purpose of this study was to assess the clinical agreement of breast radiologists with a new AI system developed to assess breast density on digital breast tomosynthesis mammograms.

Materials and Methods:
After FDA approval, starting in September 2021, 11 dedicated breast radiologists interpreting screening mammograms were presented with an automatically displayed window of breast density (categories A, B, C, D) from an AI platform integrated in our PACS system (Visage Imaging, Inc. San Diego, CA). The density output of each breast (left and right) was presented along with the option of a single combined density reading for both breasts. The radiologists were presented with two buttons to either agree or disagree with the AI’s assessment. If they disagreed, the radiologist could then click on their choice of density category. All data were recorded digitally and retrospectively tabulated for review. Cases of disagreement between radiologist and AI were assessed.

For a total of 291 days and 93,560 mammograms, the agreement of the radiologists with the AI was 99.35%. Of the 0.65% cases in which there was disagreement, the difference was only by one category. There were no cases of disagreement of more than one category. In case of radiologist disagreement, 59% cases were upgraded and 41% were downgraded. The change in density assignment for upgraded cases was: A to B (62%), B to C (30%) , C to D (8%), and for downgrading was: B to A (24%), C to B (48%), and D to C (28%). The most changed categories were upgrading A to B and downgrading C to B.

The AI tool showed very high radiologist agreement, indicating strong support for use as an adjunct or standalone density assessment tool.