Some questions about PSF_forward and Spline

[liuliuliu1111]

Hello,
I'm new to SMLM, but I'm very interested in your work because your results look amazing！Thank you so much because I have learned a lot from your code, and now I have some questions about the process psf.forward. Based on my knowledge, I think this forward function of CubicSplinePSF is used to complete the convolution work. So the first thing I want to ask is the purpose of this process. If it is so, I don't quite understand the concept of ROI defined in it. Do you cut the whole area into small pieces and realize the convolution in small regions? The second one is that _spline_impl.forward_frames seems to be encapsulated, is it used for spline interpolation of the PSF and then convolution? I don't know if it is OK to ask for the details or even the code of this part. That's all. Thank you again!

[Haydnspass]

Hey,
Thank you for your interest in our work. Answers inline:

Hello, I'm new to SMLM, but I'm very interested in your work because your results look amazing！Thank you so much because I have learned a lot from your code, and now I have some questions about the process psf.forward. Based on my knowledge, I think this forward function of CubicSplinePSF is used to complete the convolution work.

No, PSF.forward is used to calculate the PSF, i.e. the optical response of the microscope for simulation.

So the first thing I want to ask is the purpose of this process. If it is so, I don't quite understand the concept of ROI defined in it. Do you cut the whole area into small pieces and realize the convolution in small regions?

The region of interest (ROI) is an implementation detail to not calculate photons for the whole frame every time (where most of the values would be very close to zero) but only a relevant region. For splines, the ROI is typically around 30 x 30 px

The second one is that _spline_impl.forward_frames seems to be encapsulated, is it used for spline interpolation of the PSF and then convolution? I don't know if it is OK to ask for the details or even the code of this part. That's all. Thank you again!

Spline_impl forward is custom C++ / CUDA code to do the spline interpolation very fast, in pure python it would be quite slow.
Again, our splines and the Convolutions in the sense of a Convolutional Neural Network are not related.

Hope I could help.
Best,
Lucas

[liuliuliu1111]

Thanks for your reply!
I'm sorry that my description may not be clear, the convolution in my question is not the one in Convolutional Neural Network. Maybe we are talking about the same thing. What I mean is that the blurred image comes from the convolution of emitter and PSF, and it's the optical response of the microscope for simulation, right?
So do you mean that you just calculate photons around the area where there is an emitter?

[Haydnspass]

Yeah, yes, indeed psf.forward does the mathematical convolution of the psf with the emitter.

And yes the ROI is just to make computation cheaper, because the values vanish in great proximity of pixels to the emitter.