Skip to content
PyTorch Implementation of EBP - CVPR 2019
Python Shell
Branch: master
Clone or download
Latest commit f25a146 May 3, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information. init May 3, 2019
_EBP_flowchart.png init May 3, 2019 init May 3, 2019 init May 3, 2019 init May 3, 2019 init May 3, 2019
requirements.txt init May 3, 2019 init May 3, 2019 init May 3, 2019 init May 3, 2019

Elastic Boundary Projection for 3D Medical Image Segmentation

This repository is PyTorch implementation for the paper:

Elastic Boundary Projection for 3D Medical Image Segmentation
Tianwei Ni1, Lingxi Xie2,3, Huangjie Zheng4, Elliot K. Fishman5, Alan L. Yuille2
1Peking University, 2Johns Hopkins University, 3Noah’s Ark Lab, Huawei Inc., 4Shanghai Jiao Tong University, 5Johns Hopkins Medical Institute
In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 2019.


pip install -r requirements.txt # to install the packages above
  • Disk Storage > 100G
    • The generated dataset is very big due to large numbers of pivots. But the data generation process is pretty fast :)
  • Testing Time > 1h per case
    • 3D reconstruction, especially KDE and voxelization, is very slow. We will try to accelerate it in future work.


1. Download Spleen Dataset

Due to the privacy of our medical dataset, here we provide a public dataset MSD for demonstration.

Please download and unzip the spleen part Task09_Spleen.tar of MSD dataset into your $data_path (please set it in There are 41 training cases of spleen CT scans (the 20 testing cases are NOT open to the public). The shape of case is (512,512,31~168). For MSD Spleen dataset, we simply split it in half, say the first 20 for testing and the last 21 for training. You can try other divisions of the dataset.

Set the switch ENABLE_DATA_DOWNLOAD=1 in, and run it sh to reformat and split the dataset.

2. Data Generation

Set the switch ENABLE_DATA_GENERATION=1 in, and run it to generate EBP's dataset. Make sure that your disk has enough space (> 100G).

Details: according to the paper, for each pivot of each case, we will generate 2D (image, label) pairs with (HEIGHT, WIDTH) shape through ITER_TH iterations. In practice, will make a directory for each case, which is made up of self.pivot_num npy files named XYDaaabbbcccp{0,1,2,3}.npy of (ITER_TH, self.slices + IN_SLICE + 2, HEIGHT, WIDTH) = (10, 12, 120, 120) shape. In addition, it will also produce and save the important 3D matrix $C$ (self.relabel) which has the same shape as each case.
Notation: D,X,Y stand for $r,I,O$ respectively in the paper. (aaa, bbb, ccc) is the coordinate of pivot, and {0,1,2,3} is the kind of pivot defined in the program.
For more information, please see the Overall Framework (3.2), Data Preparation (3.3), and Data Generation (3.4) in the paper.

3. Training

Set the switch ENABLE_TRAINING=1 in, and run it to train and validate the network. It costs about 12 GPU hours (multi-GPU will cause long IO latency).

The decription of the 2D neural network (U/V-Net-like) can be seen in Experiments (4.1). will print training logs of the loss matrix (4, ITER_TH) per period iterations (here 4 means four kinds of pivots defined in
As for criterion, we use Online Hard Example Mining (OHEM) for training and MSE for validation. We save the model snapshots of each epoch into $data_path/models/dataset_organ$organ_id/vnetg*.pkl and normally we choose the last one to test.

4. Testing

Set the switch ENABLE_TESTING=1 and TRAINING_TIMESTAMP to refer to the snapshot model in

Testing stage is a little complicated and pretty time-consuming because it involves the following steps:

  • Iterations:
  • 3D reconstruction: which calls

In test_iter.save_iteration_results(), we save the binary matrix and other stuff for each case. Normally, test_util.mesh3d() costs much time.
For more information, please see the Testing: Iteration and Inference (3.5) and 3D Reconstruction (3.6) in the paper.


There may exist some minor typos in the codes because they are cleaned from original codes without complete tests. Welcome to open issues if you have any questions.


If this work is useful for your research, please cite our paper:

  title={Elastic Boundary Projection for 3D Medical Image Segmentation},
  author={Ni, Tianwei and Xie, Lingxi and Zheng, Huangjie and Fishman, Elliot K and Yuille, Alan L},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},


This work was mainly done when the first author did a research internship at CCVL, Johns Hopkins University. We thank all the researchers at CCVL for their insightful discussions.
This paper was supported by the Lustgarten Foundation for Pancreatic Cancer Research.

You can’t perform that action at this time.