Author: Yuchao Gu

E-mail: [email protected]

Date: 2018-05-27

Description: The code is an pytorch implementation of 《Iterative Deep Learning for Network Topology Extraction》. But I have not finished it, it is still in the patch training step.

DRIVE: Digital Retinal Images for Vessel Extraction You can download raw data here. Then we need to do some processes to generate our train data:

1、Randomly choose a foreground pixel as the center pixel, then crop a patch(64*64) around the center pixel.

2、Find pixels at the boder which are foreground. In order to detect whether they are connected to the center pixel, we make a shortest path from the border to the center. If cost of this path is smaller than INF(we set background INF), we generate a Gaussian peak around this pixel.

left is patch data, middle is label and right is our connective model ouput

python patchConnectionTrain.py

tensorboard --logdir=log

This code depends on the following libraries:

• Python 3.6
• Pytorch
• TensorboardX
• Scikit-image
• PIL

 hourglass-network
│
├── retinapatch.zip  # my preprocessed train data and label
│ 
├── gyclab # my utils for image segmentation
│ 
├── log # tensorboard visiualize
│ 
├── readmeDisplay  # some images for readme
│ 
├── result # train output
│ 
├── genConnectiveData.py # my implementation to generate data for patch model
│ 
├── patchConnectionTrain.py # train code
│
└── readme.md # introduce to this project