A Segmentation Network Based on 3D U-Net for Automatic Renal Cancer Structure Segmentation in CTA Images

Abstract

The accuracy segmentation of renal cancer structure (included kidney, renal tumor, renal vein and renal artery) in computed tomography angiography (CTA) images has great clinical significance in clinical diagnosis. In this work, we designed a network architecture based on 3D U-Net and introduced the residual block into network architecture for renal cancer structure segmentation in CTA images. In the network architecture we designed, the multi-scale anisotropic convolution block, dual activation attention block and multi-scale deep supervision equipped for the better segmentation performance. We trained and validated our network in the training set, and tested our network in the opening testing set and closed testing set of KiPA22 challenge. Our method ranked the first place in the KiPA22 challenge leaderboard, and the Hausdorff Distance (HD) of kidney, renal tumor, vein and artery achieved the state-of-the-art, also Dice Similarity Coefficient (DSC) and Average Hausdorff Distance (AVD) of renal artery. According to the results in the KiPA22 challenge, our method have a better segmentation performance in CTA images.