Conditional Convolution Generative Adversarial Network for Bi-ventricle Segmentation in Cardiac MR Images

Accurate segmentation of bi-ventricle from cardiac magnetic resonance images can provide assistance in estimation of clinical parameters and disease diagnosis for doctors. In this paper, we propose an automated and concurrent bi-ventricle segmentation method. First, we obtain region of interest (ROI) extraction for original cardiac image from large size to small size. Then we employ the conditional convolution generative adversarial network (CCGAN), which takes the extracted ROI as input, to generate mask of segmentation. The discriminator competes with the generator on the condition of the mask source to optimize the segmentation result. Finally, we get the cardiac segmentation similar to the gold standard. The proposed method is trained and tested on the data from automated cardiac diagnosis challenge (ACDC 2017). Experiment result shows our method produce better evaluation metrics compared with other advanced researches and demonstrate the effectiveness.