A Balance Cascade of Deep Neural Networks for CT Renal Segmentation

Introduction: Kidney cancer is projected to be the sixth most common cancer in men and the tenth most common cancer in women in 2018 [PMID29313949]. The morphological and anatomic features of kidneys and renal tumors have been shown to correlate with important patient outcomes [1]. Automatic segmentation with deep learning offers a way to compute these features without manual effort [2], but to our knowledge, it has not yet been applied to CT renal segmentation. Annotating a dataset for this task is challenging because manual renal segmentation is a time intensive process, and true boundaries are ambiguous in the renal hilum. In this work, we present traditional computer vision techniques to facilitate reliable ground-truth creation with minimal user-input, along with a deep supervised learning algorithm for automatic renal segmentation. Our annotated dataset and TensorFlow implementation have been made available on our website at distrob.cs.umn.edu/eus18.