Automatic Kidney and Tumor Segmentation with Hybrid Hierarchical Networks

Authors
Yuan, Yading

Issue Date
2019

Publisher
University of Minnesota Libraries Publishing

Type
Article

Abstract
Automatic segmentation of kidney and its tumors is an essential but challenging step for extracting quantitative imaging biomarkers for accurate tumor detection, diagnosis, prognosis and treatment assessment. Kidney Tumor Segmentation Challenge (KiTS) provides a common platform for comparing different automatic algorithms on abdominal CT images in tasks of 1) kidney segmentation and 2) kidney tumor segmentation . We participate this challenge by developing a fully automatic framework based on deep neural networks. By observing that clinicians usually contour organs and tumors in the axial view while evaluating the contours in 3D space, we adopt a 3-step hierarchical structure with hybrid 2D and 3D models. In the first step, a simple 2D U-Net model is trained to obtain a quick but coarse segmentation of the kidney region on the entire 3D CT volume; then another 2D U-Net using residual blocks with channel-wise attention is applied to each kidney region for kidney and tumor segmentation. At last, the segmented tumor is refined by a 3D model for final tumor segmentation. Our framework was trained using the 210 challenge training cases provided by KiTS. By 5-fold evaluation, our method achieved an average Dice Similarity Coefficient (DSC) of 0.970 on kidneys and 0.756 on kidney tumors, respectively.

Identifiers
doi: 10.24926/548719.018