Segmentation of MRI Bladder Images by Deep Fully Convolutional Network

Bladder tumor tissue and bladder wall tissue of most patients with bladder cancer penetrate each other, with variable size, shape, and position. And there are more complicated noises and artifacts in the image of bladder MRI. These maked the precise segmentation of both the tumor and the bladder wall becomes a problem for diagnosis and quantitative analysis of the next treatment. In this paper, a deep convolutional network with U-net as the basic network framework has been proposed. The residual network as a sub-module replaced the ordinary convolutional layer for downsampling, and the dilated convolution has been used to extract the information of different receptive fields in the feature map by parallel branch downsampling. Aiming at the problem with poor data sets, three methods of adding Gaussian noise, adjusting brightness and anisotropic diffusion filtering have been proposed for data augmentation. The experimental results shown that the method proposed in this paper got a DSC value of 0.9058 for bladder tumor, and the DSC value of the bladder wall segmentation reached 0.9038, which can achieve a good segmentation effect.