Pyramid Residual Convolutional Neural Network based on an end-to-end model

Abstract

Deep learning(DL) based on image classification, segmentation, detection methods have been providing state-of-the-art performance in recent years. Particularly, these techniques have been successfully applied to medical image to Help doctors have a more convenient and clear understanding of the patient's physical condition, and among them, one deep learning technique, U-Net model has become one of the most popular techniques in the field of image segmentation. In this paper, we propose a Pyramidal Residual Convolutional Neural Network based on U-Net models(PRU-Net). The proposed model takes advantage of U-Net, Residual Network as well as Image Pyramid. There are several advantages of these proposed architectures for segmentation tasks. First of all, the design of residual structure helps us to train deep structure, which often means better effect. Secondly, the pyramid residual convolution layer is used for feature accumulation, which strengthens the relationship between shallow features and deep features, and can learn image information more efficiently. Third, it allows us to design better architecture with small number of network parameters for medical image segmentation. We add pyramid residual structure to the three benchmark models, U-Net, ResU-Net (Residual U-Net) and AttU-Net (Attention U-Net), and test them on two benchmark data sets: skin cancer segmentation and thyroid nodules Segmentation. Experimental results show that these models have better segmentation effect than those without pyramid residual structure. Gaussian pyramid residual structure is better than the max-pooling.