RSNA: Podium Presentation
Matt McCabe, Leeann D. Louis, Annie Ng, Jiye Kim, Christoph I. Lee, Diana S. M. Buist, Bryan Haslam, Edgar A. Wakelin
Abstract
Purpose:
Interpretive performance in screening mammograms (SMGs) varies widely resulting in variable care for patients. Integrating AI into the screening workflow could help standardize the quality of care. However, large scale studies of the impact of AI in the real world are lacking, particularly related to radiologists with different levels of specialization. Here, we assessed the impact of a multistage AI-driven workflow on the interpretive performance of General radiologists compared to Fellowship-trained breast imagers.
Materials and Methods:
An AI-driven workflow was deployed at scale across 5 community radiology practices (109 facilities, 95 radiologists), which involved a CADe/x device providing lesion bounding boxes and suspicion categories during interpretation, plus an AI-facilitated safeguard review routing potential missed findings to breast imaging experts. Performance before and after implementation of the workflow was compared between radiologists with and without fellowship training in breast or women’s imaging. SMGs from Sept 2021-May 2022 formed the Standard of care (SOC) cohort (using only non-AI CAD tools), and exams from Aug-Dec 2022 formed the AI-driven workflow cohort. Outcomes were cancer detection rate (CDR), recall rate (RR) and positive predictive value of recalls (PPV1). Generalized Estimating Equations were used to adjust outcomes for clustering within radiologist, radiologist covariates (experience years, SMG exam volume) and patient factors (demographics, previous breast cancer, time since last SMG).
Results:
A total of 577 742 SMGs were included, 211 931 were read by 35 Fellowship-trained radiologists (SOC: 133 975; AI-Driven workflow: 77 956) and 365 811 were read by 60 General radiologists (SOC: 235 272; AI-Driven workflow: 130 539). Fellowship-trained radiologists had higher CDR (4.47, 95% CI: 3.94-5.07) than General radiologists (3.76, 95% CI: 3.46-4.08) during SOC (p=0.027). The AI-Driven workflow improved General radiologist CDR (4.99, 95% CI: 4.51-5.53, p<0.001) such that Fellowship-trained and General radiologists CDR did not differ with the AI-Driven workflow (Odds Ratio [OR]=0.95, 95% CI: 0.82-1.11, p=0.53). RR increased for General radiologists (OR=1.17, 95% CI: 1.04-1.30, p=0.007) less than the CDR increase, resulting in an increase in PPV1 (OR=1.16, 95% CI: 1.02-1.31, p=0.019), indicating the increased recalls found cancers at a higher rate.
Conclusion:
The Multistage AI-Driven workflow significantly improved CDR and PPV1 for General radiologists to levels comparable with Fellowship-trained radiologists.
Clinical Relevance/Application:
An optimized AI-driven screening workflow can improve breast screening programs by providing Fellowship-trained performance to all patients.
Disclaimers:
Breast Suite comprises multiple applications including ProFound Pro, Breast Density, Safeguard Review, Risk Assessment, BAC and DeepHealth Viewer. DeepHealth Viewer is manufactured by eRAD, Inc. and distributed by DeepHealth, Inc.
Any claims made about Breast Suite may reference claims associated with its individual components.
Not all products and functionalities are commercially available in all countries. For clearance and commercial availability in your geography of functionalities listed and compatibility with other systems, please contact your account manager.