Segmentation of Right Ventricular MR Image Based on Deep Neural Network: Dilated DenseNet of Two Level Losses

Right ventricular function assessment plays a very important role in humans' health assessment. It's hard to segment right ventricle accurately and quickly because it's an irregular crescent shape and it changes a lot from one to others, especially diseased ones. In different kinds of deep convolutional neural network (DCNN) algorithms, Dilated-DenseNet is widely used because of good features processing method, but it still has the problem of overfitting in image segmentation. So in this paper, a improved Dilated-DenseNet method is proposed to segment right ventricular MR image. Comparing to some original algorithms, the result is much more precise and stable. The dice metric of train set is 0.91 and test set is 0.90, which shows that overfitting problem has been alleviated.