Abstract PD14-04: PD14-04 Whole mammogram image-based Cox regression improves 5-year breast cancer prediction

Abstract

Abstract To efficiently capture data from mammographic breast images and classify long-term risk of breast cancer, we developed FLIP, a novel Cox regression-based framework that fully utilizes data in the mammograms beyond current density measures. FLIP use the extensive existing data that are currently ignored in the context of breast cancer risk stratification. More than 20 studies support texture features add value to risk prediction beyond breast density. However, the entire mammogram imaging data has a high dimension of pixels (~13 million per image), greatly exceeding the number of women in a cohort. FLIP was fitted and cross-validated within the Joanne Knight Breast Health Cohort excluding cases diagnosed in the first 6 months of entry. The Joanne Knight Breast Health Cohort is comprised of over 10,000 women undergoing repeated mammography screening at Siteman Cancer Center and followed since 2010. All women had baseline mammogram at entry, provided a blood sample and completed a risk factor questionnaire. Mammograms are all using the same technology (Hologic). During follow-up through October 2020, we identified 246 incident breast cancer cases (pathology confirmed) and matched them to controls from the perspective cohort based on month of mammogram and age at entry. We obtained an AUC of 0.68 (SE 0.03) including the whole mammogram image, age and BI-RADS (4th edition) density category; and AUC of 0.72 (SE 0.04) by adding in BMI and menopausal status to this model. These 5-year prediction performances exceed that of well-developed models based on epidemiologic risk factors (P < 0.001). FLIP offers standard statistical solutions and removes barriers to wider clinical use without prohibitive training data and extensive computational requirements, providing a transparent workflow ensuring high reproducibility. It should be accessible anywhere mammograms are used. We conclude that using full mammogram images for breast cancer risk prediction captures additional information on breast tissue characteristics that relate to cancer risk, and improves prediction classification. This prediction algorithm can run efficiently in real time (in seconds) with processing of digital mammograms. Thus, this model can be easily implemented in mammography screening services and other clinical settings to guide real-time risk stratification to improve precision prevention of the leading cancer in women world-wide. Further analysis will quantify the value of adding other breast cancer risk factors, including polygenic risk scores. Addition of repeated mammogram images over time should further increase classification performance. This approach has the potential to improve risk classification by using data already available for the vast majority of women already having repeated screening mammograms. Schema overview of FLIP The raw images are in the form of .dcm files before entering into FLIP. After automated processing and image alignment, the two CC-views (left and right) are average between the two breasts for characterization. The inputted 2D mammograms are first characterized with bivariate splines over triangulation to preserve spatial distribution of pixels and accommodate the irregular semi-circular breast boundary. The characterization is further optimized (see Supplemental Material) which provides a unique and closed-form solution. b. A simple Cox proportional hazards model is adopted using well-established risk factors (RF), including age, breast density (BI-RADS), BMI, menopausal status, parity, family history, and history of benign breast disease. The mammogram image acts as an additional risk factor in the Cox regression accompanied with a 2D coefficient surface. All inferential procedures with Cox regression are applicable to FLIP which provides a transparent workflow ensuring high reproducibility. h_i (t) denotes the hazard function at time t for individual i, and h_0 (t) denotes the nonparametric baseline hazard function. c. Women who are diagnosed with breast cancer within the first 6 month of their mammogram date have been removed from this analysis and we focus on the 5-year risk. Discriminatory performance is assessed with AUC and validated via a 10-fold cross-validation. Citation Format: Shu Jiang, Graham A. Colditz. PD14-04 Whole mammogram image-based Cox regression improves 5-year breast cancer prediction [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr PD14-04.