Radiomics and deep learning of diffusion-weighted MRI in the diagnosis of breast cancer

Qiyuan Hu, Heather M. Whitney, Alexandra Edwards, John Papaioannou, Maryellen L. Giger

Abstract: The incorporation of diffusion-weighted imaging (DWI) in breast magnetic resonance imaging (MRI) has shown potential in improving the accuracy of breast cancer diagnosis. Since DWI measures possibly complementary biological properties to dynamic contrast-enhanced (DCE) MRI parameters, DWI computer-aided diagnosis (CADx) can potentially improve the performance of current CADx systems in distinguishing between benign and malignant breast lesions. This study was performed on a database of 397 diffusion-weighted breast MR images (69 benign and 328 malignant). Lesions were automatically segmented using a fuzzy C-means method. The apparent diffusion coefficient (ADC)-based radiomic features were extracted and used to train a classifier. Another classifier was trained on convolutional neural network (CNN)-based features extracted by a pre-trained VGG19 network. The outputs from these two classifiers were fused by averaging the posterior probability of malignancy for each case to construct a fusion classifier. The performance evaluation for the three proposed classifiers was performed with five-fold cross-validation. The area under the receiver operating characteristic curve (AUC) was 0.68 (se = 0.04) for the ADC-based classifier, 0.74 (se = 0.03) for the CNN-based classifier, and 0.76 (se = 0.03) for the fusion classifier. The fusion classifier performed significantly better than the ADC-based classifier (݌ = 0.013). The CNN-based classifier failed to show statistically significant performance difference from the ADC-based classifier or the fusion classifier. The findings demonstrate promising performance of the proposed classifiers and the potential for DWI CADx as well as for the development of multiparametric CADx that incorporates information from both DWI and DCE-MRI in breast lesion classification.