2021 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI) | 2021
Generative Adversarial Network with Local Discriminator for Synthesizing Breast Contrast-Enhanced MRI
Abstract
Breast magnetic resonance imaging (MRI) has been widely used as a sensitive imaging technique capable of differentiating benign from malignant tumors or predicting response to treatment. Dynamic contrast-enhanced (DCE) MRI, obtained with contrast agent (CA) injection, provides hemodynamic information and tumor characteristics. However, gadolinium-based CA could deposit in the brain and other organs and have been reported to cause nephrogenic systemic fibrosis. Thus, synthetic images having similar quality compared with contrast-enhanced (CE) MRI has important clinical implications. In this study, we proposed a generative adversarial network with a local discriminator to synthesize T1-weighted CE MRI from non-contrast MRI. Residual learning and spectral normalized weights were adopted for stable convergent properties. The attention layer and local discriminator learned nonlinear mappings and enhanced the representation of the tumors better than other methods. Our approach showed better performance using four evaluation metrics both in whole breast and tumor regions. Our method would be helpful in patients who cannot undergo CE MRI and in screening for high-risk patients by substituting CE MRI.