fdk compare with iterative

[zeminLi66]

Hello, I’d like to discuss an observation with you. I had always assumed that iterative reconstruction was superior to FDK in CT imaging. However, after recently implementing an FDK pipeline, I was surprised to find that when the projection data is sufficiently complete (fully sampled), FDK performs no worse than iterative methods—in fact, it actually seems better. The image intensity in FDK is more uniform (partly due to filtering out low-frequency components). Furthermore, in industrial CT inspection, FDK yields better sharpness and contrast for defect features. The only notable drawback is that it exhibits relatively higher noise levels." (Both the iterative method and FDK process data on a projection-by-projection basis, with the iterative algorithm adopting the SART strategy.)

[zeminLi66]

On the left is FDK, and on the right is iterative reconstruction. I'm a bit confused right now: FDK seems good enough already, and it clearly has advantages in both speed and GPU memory usage. Why do we still need to research iterative methods? Or are iterative approaches truly superior to FDK, and my current setup just hasn't brought out their full potential?

[AnderBiguri]

@zeminLi66 you are right. When you have unlimited scanning and exposure time, FDK is a great algorithm, which is why its the norm in industrial CT.
This is however simply not true always. Sometimes there are constraints on acquisition times or angular sampling.
For example, in the image here: https://github.com/CERN/TIGRE/blob/master/Frontispiece/Gallery.md#conventional-micro-ct-1
If using 48h of scanning time, FDK can be as good as the iterative. But once reduced from 48h of scanning, the FDK starts to deteriorate. In the images, you can see reconstruction at approximately 18h of scanning time (subsampled), where iteratives already are better.
Many industrial applications may require scanning of all samples, or generally faster scanning. Hardware advances may lead to that, but at some point you will pass the boundary where FDK does not do that great, and you need better math.
Bu tit is true that this is not always true, and often FDK is enough.

Note however in your 2 images, you are not displaying the images in the same color-range. That is much better for comparison, as iteratives and FDK produce different histograms, and tus images. E.g. in your right image, segmenting where the chip is is easier, as its has a clearly different background color than the band in the middle, which is not true in FDK. Other examples of FDK failing can be seen in various papers in my past. E.g. the high pass filter effect near attenuating objects can be a serious disadvantage for porosity calculation: https://www.sciencedirect.com/science/article/pii/S0032386119305105