shiyuangogogo / WSISPDR

Weakly Supervised Cell Instance Segmentation by Propagating from Detection Response, in MICCAI2019.

Weakly Supervised Cell Instance Segmentation
by Propagating from Detection Response

by Kazuya Nishimura, Ker Dai Fei Elmer, Ryoma Bise

[Home] [Project] [Paper]

Prerequisites

• python >= 3.6
• ubuntu 18.04
• CPU or GPU(NVIDIA Driver >= 430)
• matlab

Installation

Python setting

Conda user

conda env create -f=env_name.yml
conda activate pytorch

Docker user

docker build ./docker
sh run_docker.sh

Graph-cut installation

Graph-cut setting

We use following code.

https://jp.mathworks.com/matlabcentral/fileexchange/38555-kernel-graph-cut-image-segmentation

mkdir graphcut 
cd graphcut
wget http://www.wisdom.weizmann.ac.il/~bagon/matlab_code/GCmex1.9.tar.gz
tar -zxvf GCmex1.9.tar.gz
matlab -nodesktop -nosplash -r 'compile_gc; exit'
cd ..

Demo

This demo is only one image's demo. If you want to apply this method to your dataset, you should prepare the likelihood map.

Back propagate from each cell

Use cuda

python propagate_main.py -g

Use cpu

python detection_train.py 

Optins:

-i :input path(str)

-o :output path(str)

-w :weight path want to load

-g :whether use CUDA

Graph-cut

matlab -nodesktop -nosplash -r 'graphcut; exit'

This is a sample code.

We don't provide dataset.

If you want to apply your dataset, you should prepare the original image and point level annotation(cell centroid). The attached text file (sample_cell_position.txt) contains a cell position(frame,x,y) as each row. Prepare the same format text file for your dataset.

Generate likelyfood map

Set the variance to a value sufficiently larger than the target object. The guided backpropagation depends on variance size.

python likelymapgen.py 

Option:

-i :txt_file_path (str)

-o :output_path (str)

-w :width (int)

-h :height (int)

-g :gaussian variance size (int)

Train

Use cuda

python detection_train.py -g

Use cpu

python detection_train.py 

Optins:

-t :train path(str)

-v :validation path(str)

-w :save path of weight(str)

-g :whether use CUDA

-b :batch size (default is 16)

-e :epochs (default is 500)

-l :learning rate(default is 1e-3)

Predict

Use cuda

python detection_train.py -g

Use cpu

python detection_train.py 

Optins:

-i :input path(str)

-o :output path(str)

-w :weight path want to load

-g :whether use CUDA

citation

If you find the code useful for your research, please cite:

@inproceedings{nishimura2019weakly,
  title={Weakly Supervised Cell Instance Segmentation by Propagating from Detection Response},
  author={Nishimura, Kazuya and Bise, Ryoma and others},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={649--657},
  year={2019},
  organization={Springer}
}