Using Uncertainty Information for Kidney Tumor Segmentation

Joffrey Michaud, Tewodros Weldebirhan Arega, Stéphanie Bricq

Abstract: Kidney cancer occurrence increases since 1990’s and its main treatment is surgery. According to this, performing automatic segmentation is an important tool to develop. In this paper, we used a two stages pipeline to get the segmentation of kidney, tumor and cyst. The first stage is used to segment the kidney region to allow us to crop the data. The second stage leverages uncertainty using Monte-Carlo dropout during training by introducing an uncertainty estimate term in the loss function.