DCFD-Net for Segmentation on MR Images of Right Ventricle

With the successful applications in image processing, convolutional neural networks (CNNs) have been transferred into medical image processing and a great success has been achieved over the past few years. The U-net architecture is one of the most famous CNN architectures for semantic segmentation in the field of medical image segmentation. However, high-level semantic feature information learned by convolution layers can't make the pixels on target image's edge easily distinguished in semantic segmentation. In this paper, we propose a novel neural network called dense connection of fully dilated-convolution network to deal with this challenging problem. The proposed DCFD-Net fuses different sizes of receptive fields through dense connection to reuse all feature maps and replace the basic architecture of down-sampling and up-sampling with dilated convolution. Two-level losses method and post processing are proposed in this paper to make model training more adequate. Our proposed neural network architecture outperforms U-Net in segmentation on MR images of right ventricle.