Semantic Segmentation of Kidney, Cyst and Tumor

MINGYANG ZANG, Artur Wysoczanski, Elsa Angelini, Andrew F Laine

Keywords: 3D U-Net, Medical Image Segmentation

Abstract: The accurate, automated detection and segmentation of re-nal tumors is of great interest for the imaging-based diagnosis, histologicsubtyping, and management of suspected renal malignancy. The KiTS21Grand Challenge provides 300 contrast enhanced CT images with kidney,tumors and cysts with corresponding manual annotation, to facilitatethe development of robust segmentation algorithms for this task. In thiswork, we present an adaptation of the historically-successful 3D U-Netarchitecture, combined with deep supervision, foreground oversamplingand large-scale image context, and trained on the majority-predictionsegmentation masks. We achieve validation performance of 96.3%, 85.6%,and 83.5% volumetric Dice score, and 91.9%, 74.9% and 72.9% surfaceDice score, on combined foreground, renal masses, and renal tumors,respectively.