PRDNet: Medical image segmentation based on parallel residual and dilated network

Abstract

This paper presents an improved semantic segmentation model for medical images named PRDNet. Common semantic segmentation algorithms only feed the features of the last layer of convolutional neural network into the decoder, which results in the position information of the first several layers of convolutional neural network is not utilized. While in our work, ResNet and dilated convolution are simultaneously used to extract multi-layer features of medical images in parallel. In the decoding stage, the multi-layer features are fused according to the structure of feature pyramid network. Moreover, several classic semantic segmentation algorithms were selected for comparison, such as UNet, Attention UNet, FPN, Deeplab v3, SENet and DANet. After a series of experiments on CHAOS and ISIC2017 datasets, the algorithm proposed by us has a 1%–4% improvement in different evaluation metrics compared with other algorithms.