2021年 · 第15卷第06期

Mammography is a commonly used but imperfect method to assess breast cancer risk. The current U.S. breast cancer screening guidelines also use breast cancer risk prediction models to guide clinical practice. The improved risk model makes the target screening strategy can detect breast cancer earlier than the current guidelines and reduce screening harm. In order to introduce the mammography-based AI deep learning risk model into clinical practice, it is necessary to further improve its accuracy and verify its effectiveness in different populations. Researchers from the Department of Electrical Engineering and Computer Science of Harvard University, MIT, United States, Imaging Physiology and Tumor Pathology of Karolinska Institutet in Sweden and the Department of Medical Imaging and Intervention of Chang Gung Memorial Hospital in Taiwan Province used 210 819 and 25 644 continuous mammography screening data from Massachusetts General Hospital (MGH), respectively, and used artificial intelligence machine learning algorithm to establish and verify the Mirai model to predict the future risk of breast cancer. Then, the continuous mammography data from MGH, Karolinska University Hospital (KUN) and Chang Gung Memorial Hospital in Taiwan Province (CGMH) were used for testing. The results showed that in the MGH test set, the C index and the area under the 5-year predicted ROC curve (AUC) of the Mirai model were 0.76 (95%CI: 0.74-0.80) and 0.76 (95%CI: 0.73~0.80), both better than the hybrid deep learning model (C-index: 0.72, 95%CI：0.69~0.75； AUC: 0.72, 95%CI: 0.69~0.76), image deep learning model (C index: 0.72, 95%CI：0.69~0.75； AUC: 0.73, 95%CI: 0.70~0.77) and Tyrer-Cuzick model (C index: 0.64, 95%CI：0.60~0.67； AUC: 0.62, 95%CI：0.59~0.66）。 In the KUH and CGMH test sets, the consistency and 5-year ROC-AUC of Mirai prediction model are also higher than those of image deep learning model. In the MGH test set, the Mirai model was able to convert 41.5% (95%CI: 34.2~48.5) of patients who developed breast cancer within 5 years were identified as high-risk patients, all higher than those of the hybrid deep learning model (39.7%, 95%CI: 32.9% ~46.5%), image deep learning model (36.1%, 95%CI: 29.1% ~42.9%) and Tyrer-Cuzick model (22.9%, 95%CI：15.9%~29.6%）（P<0.001）。 For all datasets, the prediction performance of the Mirai model is better than the other prediction models mentioned above, and the accuracy of determining patients with high risk of breast cancer is also significantly higher. The results also suggest that the AI-based deep learning model based on mammography may help to develop new breast cancer screening guidelines.