search

TristaZeng / ZS-DeconvNet

54 stars 4 forks source link

How to create 2D psf？ #15

Open Xuesm opened 3 weeks ago

Xuesm commented 3 weeks ago

Hello, thank you very much for sharing your excellent work. I have successfully trained a new model using the 2D data and PSF provided by you. I am now replacing it with my own 2D WF data, but I found that the creat_psf. m provided by you created a 3D PSF. Can you tell me how you create the 2D WF PSF? And does dxy refer to pixel size in creat_psf. m? Looking forward your prompt reply!

TristaZeng commented 3 weeks ago

2D WF PSF is the axial center slice of the 3D PSF, you can set SizeZ=1. dxy is the lateral pixel size of PSF.

Xuesm commented 2 weeks ago

thanks for your reply，helped me a lot!

Xuesm commented 2 weeks ago

Hello, may I ask how dxy is calculated in the created PDF. m file you provided? I found that the values in the file are not quite what I understood. Is there any calculation relationship between it and the size of the detector focal plane pixel-size and the lens magnification? Looking forward to your reply, thank you very much! 1

TristaZeng commented 2 weeks ago

The pixel size of the output PSF will be dxy, which you specified, and usually you specify it to be the actual sampling period (calculated according to the optical system, given the detector pixel size, magnification rate, etc) of the specimen. But in practice you can specify it to be small enough values to create a finer PSF, and enter the psf dxy value during training, and the training code will automatically interpolate the PSF to match with the data dxy.

Xuesm commented 1 week ago

Thank you very much for your reply! I found that the values of dx and dy here are the same as dxy. What do dx and dy mean？ How do you set these two values? Looking forward to your reply, thanks a lot! 2

TristaZeng commented 1 week ago

dx and dy are usually the same. They are the horizontal and vertical sampling distance.

Xuesm commented 1 week ago

Got it! Thanks bro! I will try again.

Xuesm commented 1 week ago

Hello, did you find ringing artifacts in the deconvoluted results? Is it due to inaccurate PSF? The following two images are the original image and deconved_iter50000 image of my test on ZS-DeconvNet. nirOri nirDeconv

TristaZeng commented 1 week ago

Yes it seems very much like inaccurate PSF. If you are upsampling, use /2 dxy.

Xuesm commented 1 week ago

Hello, sorry to bother you again. I don't understand what you mean by using/2 dxy for upsampling, and where is the parameter reflected in Train_ZSDeconvNet_2D.py? Another issue is that I found when I define different PSF matrix size and keep other PSF parameters the same, the deconvolution effect of the trainning model is very different, should there be limitations on the size setting of PSF? Looking forward to your reply to the above questions，thanks! fig1 is the PSF parameters we set. Fig2 is deconvoluted_50000iter result with size 527x527 PSF. Fig3 is deconvoluted_50000iter result with size 851x851 PSF. Fig4 is the parameters in Train_ZSDeconvNet_2D.py we used. psf设置 psf527_527 psf851_851 trainModelParSet

TristaZeng commented 6 days ago

Please check the simulated PSF for any difference except for the size. If the size is large enough, further increasing the PSF size is just padding more zeros on the edges. If you set upsample_flag=1 in Train_ZSDeconvNet_2D.py, dx and dy should be 1/2 of the raw data lateral sampling distance. Because they are actually the lateral sampling distance of network output.

Xuesm commented 5 days ago

Thanks，I find related description about data in train_demo_2D.sh. I will try again. And I did use simulated PSF with the same parameters except for the size. I will check again. thank you very much!