Abstract IA-21: AI in an imaging center: Challenges and opportunities

Abstract

There is significant and well-founded enthusiasm for the promise of AI to improve healthcare outcomes across diverse clinical applications in oncology. However, careful analyses of the impact of AI tools in actual “real world” clinical settings are sparse in this relatively early phase of scientific discovery. The rigor applied to deep learning model development (training, testing, external validation) must be matched with equal rigor in assessing patient outcomes after implementation of deep learning tools in actual clinical settings. We can learn from our past endeavors in computer-assisted image interpretation. Tools for computer aided detection were approved by the FDA in 1998 to help radiologists find breast cancers on mammograms. However, after widespread clinical implementation (based on promising results in reader studies and reimbursement approvals), clinical trials to assess the impact of CAD on diagnostic accuracy of mammography revealed the tools did not improve human performance in findings cancers, and in many cases increased false positive readings by humans using CAD assistance. Due to severely limited healthcare resources during pandemics, and to protect patients and healthcare workers, state governments and the COVID-19 Pandemic Breast Cancer Consortium urged providers to focus cancer screening efforts on those patients at higher risk. These mandates were and continue to be necessary responses to support fair allocation of scarce resources to maximize benefits for all patients across the full spectrum of healthcare needs and acknowledge the shift in risk-benefit ratios for screening in average risk women. After screening mammography programs were closed due to the COVID-19 pandemic, we implemented an AI image-only risk model strategy at MGH to invite women at higher risk to return to screening for breast cancer. The AI model identified the top 60% of patients at risk, who were diagnosed with 90% of cancers, while the top 60% of at-risk patients based on traditional models harbored less than 50% of cancers. Our AI image-only risk model performed best in identifying women with current breast cancer, followed by strategies which used five-year traditional risk models and included all women with a personal history of breast cancer in invitations to screen. Lifetime risk models performed extremely poorly. Operating at the traditional “high risk” threshold of >20% lifetime risk, these models identified fewer than 10% of patients with cancer in the total population screened. AI-based risk models may support accurate and equitable risk-based screening. It may be possible with early intervention to reduce the widening disparities and later stage breast cancer diagnoses we are experiencing with COVID and mitigate the risk of delayed diagnoses and increased mortality from breast cancer during and after pandemics. Strategies for careful and rigorous assessment of clinical implementation are essential for AI to have the impact in health desired by all. Citation Format: Constance Lehman. AI in an imaging center: Challenges and opportunities [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr IA-21.