Design of Superpiexl U-Net Network for Medical Image Segmentation

In recent years, the superpixel methods have been widely used in the field of medical image processing and achieved good results, such as LAW, SLIC, etc. However, there were still some problems of fuzzy classification at the edge of the tissues when these methods were used to obtain superpixels. A superpixel optimization approach based on U-Net architecture was proposed in this paper. Firstly, a bilateral filtering (BF) operation was adopted to eliminate external noisy effects at the beginning of the network, and enhance the grayscale information. Then, via combining with U-Net networks, the whole model can learn the image features and output the optimized results for the superpixel map. In terms of network design, a normalization layer was embedded behind the convolution layer at each feature-scales, in order to strengthen the sensitivity of the parameters. Experimental results show that the classification accuracy in superpixel edge is significantly improved compared with the ground truth. Moreover, this method has achieved better results in precision, recall, F-measure and computational efficiency than other classic methods.