Crossbar-Net

Crossbar-Net: Regions with Convolutional Neural Network Segmentation

Created by Qian Yu, Yinhuan Shi, Jinquan Sun and Yang Gao at Nanjing University.

Introduction

This work is designed for kidney tumor segmentation, but it can be easily modified to be used in other 2D medical segmentation applications. The code runs on Matconvnet1.0-beta20. This is the CPU version and you can change it to the GPU version easily.

Citing Crossbar-Net

If you find Crossbar-Net useful, please consider citing:

@article{Qian2018crossbar, author= {Qian Yu, Yinhuan Shi, Jinquan Sun, Yang Gao,Jianbing Zhu and Yakang Dai}, title = {Crossbar-Net: A Novel Convolutional Neural Network for Kidney Tumor Segmentation in CT Images}, journal={arXiv preprint arXiv:1804.10484v2}, year= {2018} }

Sampling training patches with basic sampling strategy

1. Place the folder "kidneytumor" in the matconvnet\examples\ folder. Place the folder "kidney-baseline-simplenn" in the matconvnet\data\ folder.
2. Four mat files should be created based on your training images and ground truth, which are listed in the matconvnet\examples\kidneytumor\basic_sampling.m.
3. Run basic_sampling.m. The data_horizontal.mat and data_vertical.mat are saved in kidney-baseline-simplenn\horizontal and kidney-baseline-simplenn\vertical, respectively.

Training sub-models

If you have prepared the the training patches according to the basic_sampling strategy, that is, the data_horizont.mat and the data_vertical.mat have been created, you can run the Patch_imdb.m to create the imdb data. Then place the imdb data in kidney-baseline-simplenn\horizontal and kidney-baseline-simplenn\vertical, respectively. Finally, runing the training_submodels.m.

Testing sub-models

Runing the test_experiment.m. This program can be runned directly without any data being prepared. Six sub-models (e.g., H1, H2, H3, V1, V2, V3) are given in the kidney-baseline-simplenn\horizontal and kidney-baseline-simplenn\vertical. In order to show the visualization clearly, we only list the results of H1 and V1, but you can test the remaining sub-models at any time in the similar manner with H1 and V1.