bala93 / multi-task-deep-network

Usage of Multi-task deep learning network for semantic segmentation in medical images

Multi-task-deep-network

Multi-task deep learning based approaches for semantic segmentation in medical images

Psi-Net: Shape and boundary aware joint multi-task deep network for medical image segmentation (EMBC 2019) Conv-MCD: A Plug-and-Play Multi-task Module for Medical Image Segmentation (MICCAIW - MLMI 2019)

Dependencies

Packages

• PyTorch
• TensorboardX
• OpenCV
• numpy
• tqdm

An exhaustive list of packages used could be found in the requirements.txt file. Install the same using the following command:

 conda create --name <env> --file requirements.txt

Preprocessing

Contour and Distance Maps are pre-computed and can be obtained from binary mask. Sample matlab codes can be found here:

• Contour: https://in.mathworks.com/help/images/ref/bwperim.html
• Distance: https://in.mathworks.com/help/images/ref/bwdist.html

Directory Structure

Train and Test folders should contain the following structure:

├── contour
    |-- 1.png
    |-- 2.png
    ...
├── dist_contour
    |--1.mat 
    |--2.mat
    ...
├── dist_mask
    |-- 1.mat
    |-- 2.mat
    ...
├── dist_signed
    |-- 1.mat
    |-- 2.mat
    ...
├── image
    |-- 1.jpg
    |-- 2.jpg
    ...
└── mask
    |-- 1.png
    |-- 2.png
    ...

Train code

UNET

base_path=''
train_path=${base_path}'/train/image' 
val_path=${base_path}'/test/image' 
model_type='unet'
object_type='polyp'
save_path=${base_path}'/models'
python train.py --train_path ${train_path} 
                --val_path ${val_path} 
                --model_type ${model_type} 
                --object_type ${object_type} 
                --save_path ${save_path}

DCAN

base_path=''
train_path=${base_path}'/train/image' 
val_path=${base_path}'/test/image' 
model_type='convmcd'
object_type='dcan'
save_path=${base_path}'/models'
python train.py --train_path ${train_path} 
                --val_path ${val_path} 
                --model_type ${model_type} 
                --object_type ${object_type} 
                --save_path ${save_path}

DMTN

base_path=''
train_path=${base_path}'/train/image' 
val_path=${base_path}'/test/image' 
model_type='dmtn'
object_type='polyp'
save_path=${base_path}'/models'
python train.py --train_path ${train_path} 
                --val_path ${val_path} 
                --model_type ${model_type} 
                --object_type ${object_type} 
                --save_path ${save_path}

Psi-Net

base_path=''
train_path=${base_path}'/train/image' 
val_path=${base_path}'/test/image' 
model_type='psinet'
object_type='polyp'
save_path=${base_path}'/models'
python train.py --train_path ${train_path} 
                --val_path ${val_path} 
                --model_type ${model_type} 
                --object_type ${object_type} 
                --save_path ${save_path}

Conv-MCD

base_path=''
train_path=${base_path}'/train/image' 
val_path=${base_path}'/test/image' 
model_type='convmcd'
object_type='polyp'
save_path=${base_path}'/models'
python train.py --train_path ${train_path} 
                --val_path ${val_path} 
                --model_type ${model_type} 
                --object_type ${object_type} 
                --save_path ${save_path}

Citations

If you use the Conv-MCD or Psi-Net code in your research, please consider citing the respective paper:

@article{Murugesan2019PsiNetSA,
  title={Psi-Net: Shape and boundary aware joint multi-task deep network for medical image segmentation},
  author={Balamurali Murugesan and Kaushik Sarveswaran and Sharath M. Shankaranarayana and Keerthi Ram and Mohanasankar Sivaprakasam},
  journal={ArXiv},
  year={2019},
  volume={abs/1902.04099}
}

@misc{murugesan2019convmcd,
    title={Conv-MCD: A Plug-and-Play Multi-task Module for Medical Image Segmentation},
    author={Balamurali Murugesan and Kaushik Sarveswaran and Sharath M Shankaranarayana and Keerthi Ram and Jayaraj Joseph and Mohanasankar Sivaprakasam},
    year={2019},
    eprint={1908.05311},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

multi-task-deep-network's Issues

Is there something to add to the verification code？

Is there something to add to the verification code？

Can you tell me what's included in the validation dataset？

Can you tell me what's included in the validation dataset？

Codes to be added

Validation and contour, distance map computations.

signed distance transform

How is the signed distance transform implemented.

训练数据从哪里下载？

您好，请问你们的训练数据从哪里下载？

Issue in Dataset preparation

During dataset preparation, I encountered challenges with calculating dist_mask, dist_contour, and dist_signed. The computation of these values is performed using the Matlab library, as indicated in the GitHub repository. I successfully obtained the contour and distance map in PNG format. Can you assist me in getting dist_mask, dist_contour, and dist_signed in a mat file?

jaccard_weight ?

what other values can jaccard_weight have other than the default 0? I mean what should be the range of it?
Thanks

Loss function and argparse

Move the loss function definitions to losses.py, add arg parse for the variables.

Contour and dist_contour

Could you tell me more detail about Contour and dist_contour file. I dont understand what does it contain? Thank you

The dataset!

Hello, where do you download your training data from?

Thanks.