Extending nnU-Net for the KiTS21 Challenge

Alex Markow, Lea Sophie Heinzelmann, Luca Emanuel Fioravanti, Matthias Ngo, Richard Hempel

Keywords: KiTS21 Challenge, U-Net, Medical Image Segmentation

Abstract: Tools for automatic semantic segmentation of 3D CT scans have shown significant results for kidney and tumour extraction using deep learning methods. The ability to produce reliable and accurate segmentations will help with surgical planning and improve the efficiency of malignant tumour detection. In this paper, we describe a three-stage pipeline for kidney, tumour and cyst segmentation based on the nnU-Net framework [1]. The initial stages extract the kidney regions and detect class presents. Then the last stage individually segments the detected classes in the cropped areas.