Abstract

Info

Author(s):: LI Jianing; WANG Yu; XIAO Hongbing; YAN Shanwu; SUN Mei; School of Artificial Intelligence, Beijing Technology and Business University, Beijing 100048, China

Keywords:: Keywords: U-Net CT image kidney tumor multi-scale convolution residual module

Abstract:: Abstract: Aiming at some intractable problems such as the uncertain size and location of kidney tumors, and overfitting etc. when using the traditional U-Net network to segment kidney tumor full-automatically, a fully automated segmentation method for kidney tumors based on multi-scale and residual U-Net (MSR U-Net) is proposed. Skip connections is added in the residual module to make the network converge faster, further avoiding the overfitting. Moreover, 3 different sizes of convolution kernels are used in the multi-scale convolution module to increase the receptive field, solving the problem of single feature size. The CT slices of 90 patients in the KITS19 database are used to carry out relevant verification experiments. MSR U-Net method obtains a Dice coefficient of 0.976 for kidney and 0.836 for tumor, indicating that the proposed algorithm is feasible and effective in the fully automated kidney tumor segmentation task.