Was the registration operation used when calculate error_map ?

[qiqi-lu]

Hi,
I was using your code to calculate the error map in SQUIRREL algorithm.
I wonder whether the registration operation was used when calculating error map, as I saw the interpolator object was not used at

NanoPyx/src/nanopyx/core/transform/error_map.pyx

Line 43 in a1fed57

cdef object interpolator = ShiftAndMagnify(verbose=False)

Best,
Qiqi Lu

[brunomsaraiva]

Hi @qiqi-lu,
During the error map calculations we are not performing registration beyond interpolating the reference image to the same size as the SR one. That line was mistankenly kept once we changed the interpolation from Catmull-rom interpolation to an FHT based interpolation (similar to the NanoJ implementation).
You can see the usage of the new one in:

NanoPyx/src/nanopyx/core/transform/error_map.pyx

Line 52 in a1fed57

result = dht_interpolation(np.asarray([img_ref]), magnification)

.

It's referencing the function implemented here:

NanoPyx/src/nanopyx/core/transform/_interpolation.pyx

Line 201 in a1fed57

def fht_space_interpolation(image, int magnification=2, bint doMirrorPadding=True):

Which uses numpy FFTs as we currently don't have our own FFT implementation in NanoPyx.

Best,
Bruno Saraiva

[qiqi-lu]

Thank you very much for you quick reply @brunomsaraiva !

I noticed that a shift was calculated in the original SQUIRREL algorithm (Supplementary Note 3 in the Supplementary Information file in the paper "Culley, S. et al. Quantitative mapping and minimization of super-resolution optical imaging artifacts. Nat Methods 15, 263–266 (2018)")

Do I need to register the SR and reference images used my own code before image intensity rescaling?

Best,
Qiqi Lu

[brunomsaraiva]

You're absolutely right, we did not implement at the time, as in our use cases we noticed most of the time we wouldn't need to register the image. But for consistency sake, I'll look into implementing the registration as part of the next release of NanoPyx (will keep the issue open until then :) )

In the meantime, if you want to perform registration before calculating the error map you can:

1. interpolate the reference image first using the fht_interpolation
2. register the image using our drift_alignment method

   NanoPyx/src/nanopyx/methods/drift_alignment/__init__.py

   Line 6 in a1fed57

   def estimate_drift_alignment(image_array, save_as_npy=True, save_drift_table_path=None, roi=None, **kwargs):

   Use apply=True , to return the aligned image
3. either crop or apply a zero mask on the SR mask on the same coordinates that were set to 0 on the reference image (due to alignment)
4. compute the error map using the interpolated and aligned reference image and the SR image with the same zero borders and using magnification=1

I haven't tested this yet, so let me know if you find any hiccups, but in principle this should work as a one-to-one of what is described in the original SQUIRREL paper.

Best,
Bruno

[qiqi-lu]

Thank you for your kind suggestion @brunomsaraiva ! I will look into it.