neel-dey / atlas-gan Goto Github PK

Hi @neel-dey ,
Sorry to bother you again. I tried to reproduce some code using PyTorch before, but I encountered issues with the generation network part. It consistently behaves differently from what I expected. Could you please provide a detailed explanation of the structure of the generation network part?Alternatively, could you help me take a look at my code to identify any issues?

Hi @neel-dey ,

First of all, thanks for making the code available.Could you please share your testing code with me, or alternatively, could you send it to my email? My email address is [email protected]. Your assistance is greatly appreciated!

Thank you sincerely.

Hi @neel-dey,

First of all, thanks for making the code available. I tried out the model on my data set and everything turned out to be great. Except that during training I noticed that starting from around 10,000 iteration, strange-looking patches of noise started to appear.

I saved the volume as .nii file and observed that there seemed to be more than one patch.

I was trying to figure out the problem on my own. Here's what I have done. I confirmed that my rough template to start with did not have strange patches. All my data were normalized to 0 to 1.
I can't figure out what is wrong. Intuitively, the presence of the regularizer should be able to prevent voxels from walking that far to the corners. Plus, similarity loss should be able to penalize the generation of random patches of noise at places where they shouldn't be. I have noticed that in the script that clip_bckgnd = False. I wonder if setting this to True could help resolve this problem.

I would really appreciate if you could provide me with some insight on how this happened. Hope to hear back from you soon!

Hi, thanks for providing such an excellent work. Could you provide your preprocessing dataset?

Hello - Thanks for sharing your work !
I would like to use your algorithm on prostate dataset and ask couple questions.
As far as I analyzed the code I do not see it taking into account spacing, orientation and direction - so I suppose all should be the same before starting training?

I have t2 weighted images in transverse, saggital and coronal plane and 3 resolutions for each 256x256; 512x512 and 1024x1024, should I train using all of the views of t2 or separately each? what do you think best do with differences in resolution - pad smaller images? resize and interpolate? train separate conditional template for each resolution?

I understand that some is unknown and possible to establish only by experimentation - but you know your tool best so any comment would be highly valuable .

Thank you !