Artificial Intelligence Risk Model (Mirai) Delivers Robust Generalization and Outperforms Tyrer-Cuzick Guidelines in Breast Cancer Screening

Abstract

Recently, in the companion to this article, Yala et al 1 reported on the validation of an artificial intelligence (AI)-aided breast cancer risk model called Mirai on the basis of traditional mammograms across multiple international cohorts. To the best of our knowledge, this is the first study suggesting that the AI-based breast cancer screening model developed on a largescale data set across seven hospitals with at least five races can deliver impressively robust transferability. Such extensive cross-regional validation is challenging because of interethnic heterogeneity of breast cancer, device bias from different institutions, and completeness of patient clinical data. 2 To overcome these challenges, this study developed a risk factor predictor that could produce risk factors referred in the Tyrer-Cuzick (TC) guidelines and then combined these factors with mammographic images. This study also leveraged a conditional-adversarial training scheme to eliminate the bias caused by different mammography machines. As a result, Mirai maintained high performance (range, 0.75-0.84) and could accurately identify high-risk subsets across all test cohorts. Moreover, this AI model shows better performance than the widely used traditional TC model in improving early detection and reducing overtreatment of breast cancer. Thus, this study concluded that Mirai has the potential to replace current breast cancer risk assessment models recommended by clinical guidelines for magnetic resonance imaging screening.

Reply to M. Eriksson et al and Z. Jin et al

“…We agree with Jin et al 2 that additional analyses are important for understanding the impact of Mirai. For this reason, we provide the models 4 to enable other researchers to conduct their own analysis.…”

“…We are eager to engage the community on this work and invite researchers to explore Mirai, which is freely available for download under an MIT license with extensive documentation. 4 In the following sections, we address some of the specific questions raised by Eriksson et al 1 and Jin et al 2 Eriksson et al 1 highlighted several points for clarification regarding the statistics in Table A4, the exclusion of 6-month interval cancers, and the best use of Mirai's risk scores.They asserted that Table A4 does not reflect the natural distribution of cases across years; however, our study did in fact leverage a representative sample of patients from each institution. The demographics for each institution are summarized in Tables 1 and A1-A3.…”

“…Instead, Mirai predicts the risk of a cancer diagnosis within 5 years.Jin et al 2 suggested various relevant measures of performance.They inquired about Mirai's performance on dense and nondense breasts. As summarized in Table S4 in Yala et al, 7 Mirai obtains consistent performance across both dense and nondense patients in the Massachusetts General Hospital testing set.They also suggested additional analyses, given Mirai's robust performance.We agree with Jin et al 2 that additional analyses are important for understanding the impact of Mirai. For this reason, we provide the models 4 to enable other researchers to conduct their own analysis.Both letters ask how Mirai was calibrated.…”

“…We thank Eriksson et al 1 and Jin et al 2 for their interest in our recent study 3 validating Mirai, a breast cancer risk prediction model, across seven institutions spanning five countries. We are eager to engage the community on this work and invite researchers to explore Mirai, which is freely available for download under an MIT license with extensive documentation.…”