Key points are not available for this paper at this time.
10519 Background: There is increasing interest in early detection of breast cancer by utilizing MRI in high-risk populations. However, it is still challenging to define and enrich the high-risk population. In this study, we developed an artificial intelligence (AI)-powered Imaging Biomarker in Mammography (IBM) to discover unique mammographic patterns, beyond simple density evaluations, that are related to breast cancer. Methods: A total of 49,577 mammography exams were collected to develop the AI-powered IBM, in which 6,218 were cancers. First, we evaluated the hypothesis that the unaffected breast of cancer patients would have a different pattern than that of non-cancer patients, by training AI (IBM-A) with unaffected breast in cancer patients and breasts of non-cancer patients. We then utilized further images of the cancer patients to train AI (IBM-B). This time we used both affected and unaffected breasts of cancer patients and breasts of non-cancer patients, allowing IBM-B to additionally learn patterns related to breast cancer. The IBMs were evaluated using the internal data (n = 2,058) that included 719 cancers. To demonstrate the feasibility of early detection by using IBM-B, it was tested with external data (n = 4,158) from an independent institution. This included pre-index exams (n = 292) taken prior to index exams acquired at the time of cancer diagnosis. Results: With the internal data, IBM-A showed AUC of 0.842, suggesting that AI could learn the difference between the normal breast of cancer patients and non-cancer patients. With IBM-B, which used additional cancer images to train, AUC was improved to 0.852. Based on the internal validation, IBM-B was chosen for the external validation, in which pre-index examinations were used only. IBM-B showed AUC of 0.777 in discriminating the pre-index exams of cancer patients and those of non-cancer patients. The radiologists excluded the apparent missed cancers (n = 87) by reviewing the pre-index exams retrospectively. After, the recalculated AUC of IBM-B was 0.770, suggesting that IBM-B can distinguish between mammograms of patients who will develop breast cancer in the future and those who will not. The mean IBM-B scores in pre-index exams of cancer group (0.580) were significantly higher than those in the normal (0.258, P < 0.001) and benign (0.258, P < 0.001) groups. Conclusions: AI-powered IBM could detect the unique parenchymal pattern associated with high breast cancer risks, and we show the potential of the AI-powered IBM to be used as an independent biomarker to select high-risk populations based on mammography alone.Table: see text
Kim et al. (Thu,) studied this question.