Depth Invariant 3D-CU-Net Model with Completely Connected Dense Skip Networks for MRI Kidney Tumor Segmentation

Abstract

Due to the impact and importance of the kidney objects in human body, the kidney tumor analysis from three dimensional CT and MRI medical images becomes a pivotal research topic, which helps in diagnosing the kidney diseases like kidney stones, polycystic and kidney tumors etc. In deep learning, U-Net became a prominent and reliable solution for kidney image analysis and objects segmentation process. Although several research works were focused on kidney object detection and tumor segmentation from medical images, they are suffering from some intrinsic limitations due to: variance in network depths, enforced feature fusion, segmentation errors and inaccuracy. In order to address these limitations in kidney tumor segmentation process, in this paper we proposed the 3D-CU-Net model for kidney tumor segmentation, which is a custom variant of the U-Net. In 3D-CU-Net, the encoder-decoder network model is unified to tolerate the depth invariance issues, while training various input images with the same model. Completely connected dense skip connections are designed at each layer of 3D-CU-Net, to control the enforced feature fusion and to extract the crucial features. An integrated loss function is designed with Binary Cross Entropy (BCE) and Soft-Dice Coefficient (SDC) to mitigate the segmentation errors and inaccuracy. Experiments on TCGA-KIRC dataset with 3D-CU-NET recorded the high accuracy in kidney tumor segmentation with mIoU (91.21%) and mDSC (92.69%).