Multiparametric MRI deep learning model based on dynamic Contrast-enhanced and apparent diffusion coefficient map enables accurate prediction of benign and malignant breast lesions.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, Xinhua Hospital, Shanghai Jiao Tong University School of Medicine, No. 1665 Kongjiang Road, Shanghai, 200092, China.
- School of Software and Microelectronics, Peking University, Beijing, 102600, China.
- Department of Medicine, Deepwise AI Lab, Beijing, 100080, China.
- Department of Radiology, Xinhua Hospital, Shanghai Jiao Tong University School of Medicine, No. 1665 Kongjiang Road, Shanghai, 200092, China. [email protected].
- Department of Radiology, Xinhua Hospital, Shanghai Jiao Tong University School of Medicine, No. 1665 Kongjiang Road, Shanghai, 200092, China. [email protected].
- Faculty of Medical Imaging Technology, College of Health Science and Technology, Shanghai Jiao Tong University School of Medicine, Shanghai, 200025, China. [email protected].
Abstract
The study aims to develop a deep learning (DL) model based on multiparametric magnetic resonance imaging (MRI) for distinguishing between benign and malignant breast lesions. A total of 556 lesions (307 malignant, 249 benign) in 509 patients were pooled in the training/validation datasets between November 2018 and October 2019 in this retrospective study. A combined DL model based on the dynamic contrast enhanced-MRI (DCE-MRI) and apparent diffusion coefficient (ADC) map was developed to characterize breast lesions. Model performance was evaluated by using areas under the receiver operating characteristic curve (AUC) in the validation dataset and an independent testing dataset consisting of 243 lesions in 225 patients, and compared with other combined and single-parametric DL models. The predictive performance for malignancy was also compared between the DCE-ADC combined DL model and human readers. The DCE-ADC combined DL model achieved the highest diagnostic efficiency with the AUC, accuracy, sensitivity, and specificity of 0.889, 82.5%, 80.7%, and 84.1% for predicting malignant breast lesions, surpassing other combined and single-parametric DL models. The DCE-ADC combined DL model achieved good performance (accuracy:82%) and outperformed both the junior radiologists (82% vs. 70%, p = 0.073; 82% vs. 72%, p = 0.142). The diagnostic performance of two junior radiologists was improved after artificial intelligence assistance with AUCs increased to 0.798 and 0.772 from 0.689 to 0.708, respectively. The DCE-ADC combined DL model shows promising diagnostic performance and has good potential to assist junior radiologists in improving diagnostic efficacy, which can facilitate clinical decision-making. Further studies will validate these findings in prospective, larger cohorts, multicenter, multiscanner and multinational studies.