E1969. Real World Effect of Artificial Intelligence Software on Digital Breast Tomosynthesis Screening Cancer Detection, AIR Rates, and PPV
  1. Hector Diaz de Villegas; University of Florida - Jacksonville
  2. Meredith Peratikos; Biostatistics and Data Science, LLC
  3. Julie Shisler; JLS Consulting
  4. Alicia Toledano; Biostatistics Consulting, LLC
  5. Robert Nishikawa; University of Pittsburgh
  6. Emily Conant; University of Pennsylvania
  7. Haley Letter; Mayo Clinic
This study assessed the “real world” impact of an artificial intelligence (AI) tool designed to detect breast cancer in digital breast tomosynthesis (DBT) screening exams on standard outcome statistics following 12 months of utilization in an academic breast center.

Materials and Methods:
An IRB-approved retrospective study of outcome statistics with five subspecialized radiologists (breast imaging experience 3 - 35 years) across 3 academic breast imaging sites during September 1, 2020 through August 31, 2021. All mammograms were performed with DBT (Hologic Dimensions) and interpreted on Visage PACS. Cancer case characteristics were abstracted from medical records for biopsy-confirmed true positives (TP). Site 1 had AI (iCAD ProFound AI V2.0) available during the 12-month period while two locations (Sites 2 and 3) did not. Co-primary endpoints were cancer detection rate (CDR) per 1000 screened and abnormal interpretation rate (AIR). Secondary endpoints included positive predictive values (PPVs) for cancer among screenings with abnormal interpretations (PPV1) and for biopsies performed (PPV3). Odds ratios (OR) with two-sided 95% confidence intervals (CIs) summarized impact of AI across radiologists using generalized estimating equations.

Five radiologists interpreted screening DBT studies at all 3 sites including 5,883 (43 TP) with AI and 7,002 (42 TP) without AI. The CDR per 1000 screened was 7.3 with AI, 5.9 without AI (odds ratio 1.3, 95% CI: 0.9 - 1.7). AIR was 11.7% with AI, 11.8% without AI (OR 1.0, 95% CI: 0.8 - 1.3). PPV1 was 6.2% with AI, 5.0% without AI (OR 1.3, 95% CI: 0.97 - 1.7). PPV3 was 33.3% with AI, 32.0% without AI (OR 1.1, 95% CI: 0.8 - 1.5). Among cancer cases without AI, 81% were invasive, 12% were low grade, and 24% had calcifications only. Among cancer cases with AI, 65% were invasive, 23% were low grade, and 37% had calcifications only. Among invasive cancers, all invasive lesions in the examination were < 20 mm for 82% without AI versus 79% with AI. Although evidence is limited by the low numbers of total cancers in the 12-month period, results are consistent with prior reader studies. There are modest improvements in CDR and PPV1 with AI with similar AIR across groups.

Real-world interpretation of DBT with the availability of an AI detection system in clinical practice resulted in clinically relevant increases in cancer detection without increasing recalls.