Multi-objective Evolutionary Architecture Search of U-Net with Diamond Atrous Convolution

U-Net and its variants have played important roles in the field of medical image segmentation. However, U-Nets based on conventional 3 * 3 convolution still have some shortcomings, such as the lack of deformation of receptive field. In addition, due to the limited computing resources and memory space on many machines, the allowed sizes of networks deployed on them are also limited. However, it may not be effective to manually design the architectures of U-Nets. In this paper, a U-Net architecture with diamond atrous convolution (DAU-Net) is presented. Furthermore, a multi-objective neural architecture search method with channel sorting of DAU-Net is proposed to search for the better U-Net architectures. Experimental results on the ISIC 2018 dataset of melanoma segmentation show that the proposed method obtains a series of network architectures with different sizes, and the obtained architectures achieve obvious improvements in term of both model sizes and prediction accuracies compared with several popular and manually designed variants of U-Net.