Localization and direction of vessels with semantic segmentation by Unet

Table of Contents

• Result
• Installation
• Structure
• Usage
• Reference
• License

I. Result {#result}

Class	IoU
Background	0.970
Vertical	0.757
Horizontal	0.786

1. Vertical vessel

2. Horizontal vessel

II. Installation

Python >= 3.6 ,Recommend to use Anaconda

matplotlib==3.2.2
numpy==1.19.2
Pillow==8.2.0
pytorch-gpu==1.10.2
torchvision==0.4.2
tensorboard==2.6.0
future==0.18.2
tqdm==4.59.0
scikit-image==0.17.2
torchio==0.18.76

To install for Ubuntu,

$ conda install -r requirements.txt

III. Structure

unet-multiclass-pytorch/
    - checkpoints/
    - data/
    - model/
    - runs/
    - Unet/
    - utils/
    - video/
    - params.json
    - README.md
    - requirements.txt
    - train.py
    - video_pre.py

in which:

• checkpoints/ store the best models when training
• data/ contains training data and masks
• model/ contains the trained model
• runs/ contains Tensorboard summary files
• Unet/ contains U-Net structure
• utils/ contains model parts and model related functions
• video/ contains the video that video_pre.py needs
• parameters.json define all the parameters of the training and prediction
• README.md contains tutorial
• requirements.txt contains the necessary packages
• train.py is the main script for model training
• video_pre.py is the main script for video prediction

IV. Usage

Example

Parameter	Value
epoch	20
batch_size	2
learning_rate	0.1
folder	./data/2classify

1. Terminal

python train.py -e 20 -b 2 -l 0.1 -folder ./data/2classify

2. JSON file (recommend)

Mode	Explanation
train	training parameters
prediction	prediction parameters
change_label	change colors of labels

V. Reference

[1] https://github.com/wpddmcmc

[2] https://github.com/France1/unet-multiclass-pytorch

VI. License

MIT © HYM