DDU-Net: A dual dense U-structure network for medical image segmentation

Abstract

Medical image segmentation is one of the important steps in medical image analysis and has a wide range of applications and research values in medical research and practice. However, it is still a challenging task due to the characteristics of medical images with diverse lesion scales, blurred structural boundaries and numerous modalities. To address these challenges, we design a simple yet powerful dual dense u-structured network (called DDU-Net) for medical image segmentation. Specifically, we first construct dual encoders with the densely connected, where the first encoder uses DenseNet, whose backbone network is pre-trained on ImageNet, as a fixed feature extractor, and the second encoder uses a network structure similar to the first encoder. Both encoders try to encode information on the input image, and each layer is directly connected to the next layer in a feed-forward manner. This approach not only greatly reduces the number of parameters in the network, but also makes it easier to train with smaller samples. We then design up-sampling paths applicable to the dual encoder structure, and construct deeper decoders through densely connected convolutional layers, fusing the low to high level multi-scale semantic information learned by the two encoders. Finally, we employ a multiple loss function that combines boundary and content information to make DDU-Net more focus on the accuracy of boundary delineation to improve the overall segmentation performance. We have conducted comprehensive experiments on five different medical image segmentation datasets, including skin lesion segmentation, nuclei segmentation, lung segmentation, gland segmentation and vessel segmentation (IoU metrics for DDU-Net are 0.790, 0.860, 0.925, 0.812 and 0.696 respectively). Compared to baseline U-Net and other state-of-the-art methods, DDU-Net achieves competitive segmentation performance in both qualitative and quantitative evaluation.