Ensembled Multi-Stage Approach for Automated Segmentation of Kidney, Tumor, and Cysts in the Kits21 Challenge Dataset

Panagiotis Korfiatis, Adriana Gregory, Timothy Kline

Abstract: There still exists a large opportunity to develop highly accurate, fully automated algorithms for the task of segmenting not only the kidneys, but also pathological structures such as tumors and cysts. In this manuscript we present our results from developing and applying an ensembled multi-stage approach for learning the multi-class Kits21 challenge dataset. Our approach first learns the task of segmentation for the kidney, tumor and cyst. The tumor and cysts are then learned from a multi-channel approach with input from the first stage mask, as well as the input image. Lastly, tumor and cysts are separately distinguished using an unsupervised classification approach.