Study design

The primary objective of this study was to develop a model to assess breast cancer risk and to validate its performance across diverse populations and clinical settings. We hypothesized that by carefully designing the algorithm, Mirai would significantly outperform the Tyrer-Cuzick model and other deep learning models trained on the same MGH dataset, namely ImageOnly DL and Hybrid DL, in predicting future risk. Our secondary objective was to demonstrate the ability of Mirai to identify high risk cohorts, and to compare it to alternative risk models. 

To develop Mirai, we collected consecutive screening mammograms from 80,134 patients screened between January 1st, 2009 and December 31st, 2016 at Massachusetts General Hospital (MGH) under approval of MGH’s Institutional Review Board and in compliance with the Health Portability and Accountability Act. Mammograms were taken either on a Selenia Dimensions device (Hologic) or a Lorad Selenia device (Hologic). We obtained outcomes through linkage to a local five-hosptial registry in the Massachusetts General Brigham healthcare system, alongside pathology findings from MGH’s mammography electronic medical record. We excluded patients who did not have at least one year of screening followup, who were diagnosed with other cancers (e.g., sarcoma, etc.) in the breast or did not have all four views (L CC, L MLO, R CC, R MLO), to identify 70,972 patients. Patients were randomly split into 56,786 for training, 7,020 for development, and 7,166 for testing. To enable fair comparison against the Tyrer-Cuzick model, we excluded 161 patients with prior history of breast cancer from the test set, leaving 7,005 patients. Since each patient had multiple exams, this resulted in 210,819, 25,644 and 25,855 exams of training, development, and testing respectively. We refer to an exam as “positive” if it was followed by a pathology-confirmed cancer diagnosis within five-years. We collected detailed risk factors, including those used by the Tyrer-Cuzick version 8 model (TCv8), from provider- and patient-entered information in the mammography reporting system, and associated each mammogram with patient risk factors as they were present at the time of mammography. Detailed demographics are shown in Table 3, and our data collection procedure is illustrated in Figure 4. 

To evaluate the ability of Mirai to generalize to new populations, we collected the Karolinska and CGMH datasets under approval of the relevant Institutional Review Boards. The Karolinska dataset was extracted from the Cohort of Screen-Aged Women (CSAW) (56). All women aged 40-74 within the Karolinska University uptake area who had attended screening and were diagnosed with breast cancer, without implants and without prior breast cancer, from 2008-2016 were included, as well as a random sample of controls with at least two years of followup, from the same time period. The full Karolinska case-control dataset included 11,301 women, and 70% of both cases and controls were randomly selected for inclusion in this study.  We included all mammograms, acquired on Hologic machines, from 2008-2016 for the included women that contained all four views (L CC, L MLO, R CC, R MLO), resulting in 19,328 exams from 7,353 patients. To create the CGMH dataset, which consisted of 13,356 exams from 13,356 patients, we selected random women undergoing screening mammography there between 2010-2011 that were aged 45-70 or were aged 40-44 and had a family history of breast cancer. Cancer outcomes were obtained from the national cancer registry. In both datasets, we excluded patients who did not have at least one year of screening followup or did not have all four views (L CC, L MLO, R CC, R MLO).  We obtained mammographic breast density assessments for both the Karolinska and CGMH datasets using a clinically validated deep learning model trained on the MGH dataset (57, 58). More details about these datasets are available in Table 4 and Figure 4. We emphasize that the Karolinska and CGMH datasets were only used for testing.