search

simontomaskarlsson / GAN-MRI

Code repository for Frontiers article 'Generative Adversarial Networks for Image-to-Image Translation on Multi-Contrast MR Images - A Comparison of CycleGAN and UNIT'
GNU General Public License v3.0
227 stars 77 forks source link

training dataset #5

Closed kingaza closed 5 years ago

kingaza commented 5 years ago

Hello @simontomaskarlsson ,

Is it possible to share your dataset to me?

I search images from HCP project (https://ida.loni.usc.edu), but can't find the 1113 subjects as you mentioned in the paper.

There are only found a few series like:

Best Regards, Kingaza

simontomaskarlsson commented 5 years ago

Hi @kingaza, Unfortunately I can't share the dataset since it was financed and provided by the Department of Biomedical Engineering for research purposes.

The reason you don't see the dataset might be due to your access level.

Best regards, Simon

kingaza commented 5 years ago

hi Simon,

Thanks for your reply. So I have to use my own dataset. Do you have any suggestion, i.e. image size, number, etc? thanks again.

Best Regards, Kingaza

simontomaskarlsson commented 5 years ago

What I recommend is that you take a look at the original CycleGAN repository where you can find and download the datasets they have used. These are available at https://github.com/junyanz/CycleGAN#datasets

The 256x256 pixel images can be used with the Keras code at https://github.com/simontomaskarlsson/CycleGAN-Keras and with modification to the number of channels the UNIT model in this repository can be used as well.

With larger image size is is harder to get a good result. Many articles try to attack this problem in many different ways and as of now I think 256x256 is a good place to start, especially if you want to try out the Keras models in this repository. In most cases, more images gives a better result. But creating a well distributed and big dataset if of course hard, and thereby hard to find. That is why I recommend the generous free to use datasets found at the original CycleGAN repository.

Regards, Simon

kingaza commented 5 years ago

What I recommend is that you take a look at the original CycleGAN repository where you can find and download the datasets they have used. These are available at https://github.com/junyanz/CycleGAN#datasets

The 256x256 pixel images can be used with the Keras code at https://github.com/simontomaskarlsson/CycleGAN-Keras and with modification to the number of channels the UNIT model in this repository can be used as well.

With larger image size is is harder to get a good result. Many articles try to attack this problem in many different ways and as of now I think 256x256 is a good place to start, especially if you want to try out the Keras models in this repository. In most cases, more images gives a better result. But creating a well distributed and big dataset if of course hard, and thereby hard to find. That is why I recommend the generous free to use datasets found at the original CycleGAN repository.

Regards, Simon

thx @simontomaskarlsson