search

ginobilinie / medSynthesis

medical image synthesis with GAN framework based on tensorflow (we also have a better version in pytorch: https://github.com/ginobilinie/medSynthesisV1)
MIT License
72 stars 24 forks source link

trouble to understand latent space #6

Closed hnbonsou closed 5 years ago

hnbonsou commented 5 years ago

Hi all,

Please, I am struggling to understand how the latent space is used in your code. I would like to know how MRI patches are related to the latent space to synthesize the pseudo-CT. If one can consider that the distribution used to initialize the filters can be considered as the latent space.

Thank you in advance for your understanding.

ginobilinie commented 5 years ago

I'm confused what you mean, can you please specify it?

hnbonsou commented 5 years ago

I thought that the MRI patches should be somehow connected to a known distribution which is noise (commonly called the vector Z in the GAN formulation) before to be introduced in the generator. Please, how is this connection established in your approach?

hnbonsou commented 5 years ago

Please for more precision, I put it on the graph attached scheme.pptx

rogertrullo commented 5 years ago

Since we want to generate CT images with a one to one correspondance, we do not add any noise in the input, so the second graph in your ppt is more accurate

hnbonsou commented 5 years ago

Thank You a lot, I understand now.

ginobilinie commented 5 years ago

@hnbonsou Roger's explanation is exactly what we mean. In addition, if it is one-to-more reflection, we can add some noise along with the input source image.

hnbonsou commented 5 years ago

@ginobilinie Thanks a lot for this clarification. I will go through that.