RSNA: Poster Presentation
Leeann D. Louis, Annie Ng, Edgar A. Wakelin, Jiye Kim, Christoph I. Lee, Diana S. M. Buist & Bryan Haslam
Abstract
Purpose
AI can support increased breast cancer detection; however, there is no large-scale data on cancer subtypes in these additional cancers in a United States (US) screening setting. Prospective studies of AI outside the US have shown AI can help detect more invasive cancers and ductal carcinoma in situ (DCIS) at higher rates. The way AI is employed can also impact the rate and types of additional cancers detected. This study assesses the cancer subtypes detected in a large-scale deployment of a multistage AI-driven workflow compared to the standard of care (SOC).
Methods
A multi-stage AI-driven workflow was deployed across 109 sites in 4 states in the US and involved concurrent reading with a CADe/x device, followed by an AI-facilitated safeguard review routing potentially missed findings to breast imaging experts for an additional review. The SOC cohort included screens between Sept 2021 – May 2022. The AI-driven workflow cohort included screens between Aug 2022 – Dec 2022. Cancer detection outcomes included cancer detection rate (CDR). The distribution of cancer subtypes detected by SOC versus the AI-driven workflow was assessed for: cancer type, grade, invasivity, hormone receptor status (ER, PR, HER2), triple negative and luminal B proportions and those classified as ‘dangerous’ (i.e. a combination of triple negative, Luminal B, and HER2+). Distributions were compared using unadjusted chi-square tests.
Results
A comparison of SOC (n=1491) and AI (n=963) workflows in 2455 cancer cases showed similar invasive cancer rates (67.7% vs 67.5%). The majority of cancers detected with the SOC and AI-driven workflow were intermediate or high grade (80.6% SOC, 80.7% AI). Subtype distribution in SOC was Triple Neg 9.4%, Luminal B 3.7%, HER2+ 5.7%, and Dangerous 15.1%, while AI showed 9.8%, 4.8%, 7.1%, and 16.9% respectively. No differences were found in the distribution of cancer type, grade or invasiveness (p>0.05) between cohorts. The deployment of the AI-driven workflow (N = 208 891; 1 166 cancers) resulted in significantly increased CDR (5.58 vs 4.59, p<0.001) and PPV1 (5.01 vs 4.35, p<0.001) compared to SOC (N = 307 692; 1702 cancers). Accounting for the 21.6% CDR increase, an estimated 20 additional dangerous cancers were found.
Conclusion
Employing the multistage AI-driven workflow significantly improved clinical outcomes in terms of CDR in a cohort of over 2400 cancers, with the majority of cancers detected being clinically relevant and no corresponding increase in the proportion of DCIS.
Clinical Relevance/Application
Detecting more clinically relevant cancers without increasing the proportion of DCIS diagnoses shows the AI-driven workflow investigated here increases the benefits of screening mammography without increasing potential harms.
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.