Authors
Annie Ng, Leeann D Louis, Edgar A Wakelin, Matthew McCabe, Jiye Kim, Christoph Lee, Diana S. M. Buist, Bryan Haslam
Purpose
Women with dense breasts have a higher risk of developing breast cancer and a higher risk of cancer going undetected given cancer is more challenging to detect in dense breast tissue. AI has been shown to support increased cancer detection in breast screening, however, large-scale stratified analyses by breast density are required to assess the clinical impact across these relevant subgroups. This study assesses the clinical outcomes of a large-scale deployment of a multistage AI-driven workflow, including stratified analyses by breast density.
Methods
A multistage AI-driven workflow was deployed across 109 sites and four states in the United States and involved concurrent reading with a CADe/x device, followed by an AI-facilitated safeguard review routing potentially missed findings to breast imaging experts. The standard of care (SOC) cohort included screens between Sept 2021 to May 2022. The AI-driven workflow cohort included screens between Aug 2022 and Dec 2022. Cancer detection outcomes including cancer detection rate (CDR), positive predictive value (PPV1), and cancer subtype proportions (grade, invasivity, and hormone receptor status (ER, PR, HER2)) were stratified for women with dense and non-dense breasts, defined as BIRADS categories C or D and A or B, respectively. Results for women with dense and non-dense breasts were compared using an unadjusted chi-square test.
Results
The deployment of the multistage AI-driven workflow (N = 208 891) resulted in significantly increased CDR (6.18 vs 5.04, p<0.001) and PPV1 (4.25 vs 3.68, p=0.008) for women with dense breast as well as non-dense (CDR: 5.10 vs 4.21, p<0.001; PPV1: 6.09 vs 5.35, p=0.014) compared to SOC (N = 307 692). The majority of cancers detected by the AI-driven workflow were intermediate or high grade (80.6%) and invasive (67.5%). No differences were found in the distribution of cancer grade, invasiveness, and hormonal receptor status (p>0.05) for women with dense versus non-dense breasts.
Conclusion
The deployment of a multistage AI-driven workflow significantly improved clinical outcomes (CDR and PPV1) for women with dense and non-dense breasts, with the majority of cancers detected being clinically relevant.
Clinical Relevance/Application
Detecting more clinically relevant cancers in women across all breast densities demonstrates the effectiveness and generalisability of an AI-driven workflow to support equitable clinical outcomes in screening mammography.