Retaining high-frequency side of data

[amineuron]

Hi,

Thank you for developing this tool—it's been working very well, especially in handling the organell missing wedge. However, its clear that most of the high-frequency side of data seems to be lost during the denoising process. I’m wondering if there's a way to retain these finer details.

• Do you have any recommendations for preserving high-frequency information?

• Is there a way to adjust the parameters to maintain the high-frequency data while still benefiting from the denoising?

• Could reducing subtomo_overlap help in retaining details, or would that not make much of a difference due to the final averaging step with the two half-maps?

Any guidance on this would be greatly appreciated!

[SimWdm]

Hi @amineuron,

Thanks for trying DeepDeWedge! I am glad to hear that you find it useful!

I have also noticed that DeepDeWedge sometimes removes some of the high frequency components to produce cleaner tomograms.

My hypothesis is that this is because some high frequency details are "below the noise level", so the network cannot learn to produce these details based on the noisy inputs and targets it is trained on. In other words: The network sometimes has a hard time learning to distinguish between high-frequency detail and noise, and ends up removing parts of the detail to produce a cleaner-looking tomogram. Such a trade-off between cleaner-looking images and removal of high-frequency components is often observed with denoising algorithms.

Unfortunately, I don't have a very satisfactory answer as to what you can do to solve the problem. However, I think the following might help to some extent: You could check the refined tomogram every few epochs before the model fitting has converged (e.g. every 10 epochs, but the exact number depends on how fast the model fitting converges). It is possible that in these earlier epochs, the denoising is not strong enough to remove the high frequency details because the network has not yet learned to do so. However, earlier models may also be less effective at removing noise in low frequency detail and lacking wedge filling, so (assuming my hypothesis is correct) you would need to manually find a sweet spot.

I've never tried this approach, but I'd be very interested to hear if it works for you!

Best,
Simon

[SimWdm]

Hi @amineuron,

another observation relevant to your question is that, in the process of generating model inputs and targets, splitting the tilt series based on movie frames typically preserves more high frequency details during model fitting compared to splitting the tilt series into even and odd images (see Figure 4 in the DeepDeWedge paper).

Best,
Simon

[amineuron]

Hi Simon,

Thank you so much for your comprehensive and thoughtful response! I appreciate the insights you've shared.

I’ve been using the movie frames to build the odd and even observations, as you suggested. Regarding your suggestion to check the earlier epochs, I went ahead and monitored the process from epochs 400 to 1000, and didn’t notice a significant difference and at lower epochs impact whole denoising, when it comes to seeing finer structures like neurofilaments, which are visible in the normal reconstructions, still facing challenges.

In addition to this, I tried a different approach that seems to have helped, though I haven’t fully evaluated it yet. When creating the mask, instead of focusing just on high-density areas, I adjusted the parameters to sample more uniformly across the entire tomogram (modifying the density_percentage and std_percentage values). I also used a smaller stride to sample more thoroughly. These changes seem to have improved the details, but I think I still need to further optimize them to get the best results.

Thanks again for your help, I’ll keep experimenting and will share if I find the right setup for our tomograms.

Bests,
Amin