A double cascaded framework based on 3D SEAU-Net for kidney and kidney tumor Segmentation

Authors
Cheng, Jianhong
Liu, Jin
Liu, Liangliang
Pan, Yi
Wong, Jianxin

Issue Date
2019

Publisher
University of Minnesota Libraries Publishing

Type
Article

Abstract
Accurate segmentation of kidney and kidney tumor from CT-volumes is vital to many clinical endpoints, such as differential diagnosis, prognosis and radiation therapy planning. While manual segmentation is subjective and time-consuming, fully automated extraction is quite imperative and challenging due to intrinsic heterogeneity of tumor structures. To address this problem, we propose a double cascaded framework based on 3D SEAU-Net to hierarchically and successively segment the subregions of the target. This double cascaded framework is used to decompose the complex task of multi-class segmentation into two simpler binary segmentation tasks. That is to say, the region of interest (ROI) including kidney and kidney tumor is trained and extracted in the first step, and the pre-trained weights are used as the initial weights of the network that is to segment the kidney tumor in second step. Our proposed network, 3D SEAU-Net, integrates residual network, dilated convolution, squeeze-and-excitation network and attention mechanism to improve segmentation performance. To speed training and improve network generalization, we take advantage of transfer learning (i.e., weight transfer) in the whole training phase. Meanwhile, we use 3D fully connected conditional random field to refine the result in post-processing phase. Eventually, our proposed segmentation method is evaluated on KiTS 2019 dataset and experimental results achieves mean dice scores 93.51% for the whole kidney and tumor, 92.42% for kidney and 74.34% for tumor on the training data.

Identifiers
doi: 10.24926/548719.067