This document summarizes research on using deep learning to achieve isotropic super-resolution for fluorescence microscopy. It describes how the method can address anisotropy issues in confocal and light-sheet fluorescence microscopy by learning a mapping between low and high-resolution volumes without needing matched data or priors. Simulation results and experiments on real microscopy data demonstrate how the method enhances resolution isotropically and corrects artifacts in both modalities.
Call Girls in Hauz Khas Delhi 💯Call Us 🔝9953322196🔝 💯Escort.
Deep learning enables reference-free isotropic super-resolution for fluorescence microscopy
1. Deep learning enables reference-free
isotropic super-resolution
for volumetric fluorescence microscopy
Peter Park
Research Scientist
BISPL - BioImaging, Signal Processing, and Learning Lab.
Graduate School of AI
KAIST, Korea
2. Index
1. Problem formulation – how deep learning can address anisotropy in
fluorescence microscopy.
2. Proof of concept – simulations studies
3. Super-resolution in real-world data – in CFM
4. Blind deconvolution in real-world data – in LSFM
5. Artifact correction – in LSFM
CFM: confocal fluorescencemicroscopy, LSFM: lightsheetflourescencemicroscopy
3. Lateral plane: high resolution
Axial plane: low resolution
In fluorescence microscopy, how you look determines how sharp the image is.
*Mouse cortical regions imaged with confocalfluorescence microscopy
Lateral
Axial
Lateral
20 μm
4. How do we learn if we have no matching
data nor any priors?
5. Lateral plane: high resolution
Axial plane: low resolution
Learning the mapping b/w distributions -> optimal transport approach
X
Y
Z
Low Res. High Res.
G
F
Lateral plane
Axial plane
6. Implementation of the intuition
X
Y
Z
Low Res. High Res.
G
F
Lateral plane
Axial plane
X
Z
Y
Anisotropic original Isotropic reconstruction Anisotropic reconstruction
X
Z
Y
G F
Cycle-consistency loss
X
Z
Y
( , , ) ( , , )
DX
( , , )
( , , )
DY
Strategy: train G and F by training discriminators, and :
DX DY
7. Implementation of the idea
X
Z
Y
Anisotropic original Isotropic reconstruction Anisotropic reconstruction
X
Z
Y
G F
Cycle-consistency loss
X
Z
Y
( , , ) ( , , )
DX
( , , )
( , , )
DY
Strategy: train G and F by training discriminators, and :
DX DY
+
+
UNET
DLG
DLG
OR +
+
UNET
DY
(2)
DY
(1)
DX
(1)
DX
(2)
, , , :
G:
F:
PatchGAN
Network structures
11. Simulation results – large scale quantification
y x
z
y x
z
y x
z
0
1
Input Network Ground-truth
PNSR (dB) / MS-SSIM: 16.94 dB / 0.70 19.03 dB / 0.86
12. Simulation results – Across different conditions
Std. of axial Gaussian blurring Axial subsampling rate
PSNR
(dB)
MS-SSIM
PSNR
(dB)
MS-SSIM
14. How it worked? – restoration of information in frequency domain
0
255
FFT
15. How it worked? – recovering the blurring PSF
0
1
Blurring PSF
X
Z
Y
Anisotropic original Isotropic reconstruction Anisotropic reconstruction
X
Z
Y
G F
Cycle-consistency loss
X
Z
Y
𝐵𝑙𝑢𝑟𝑟𝑖𝑛𝑔 𝑃𝑆𝐹 ≅ 𝐹(𝛿 𝑥 )
0
0.4
Estimated PSF
-> Albeit with the scaling mismatch, 3D geometry of the blurring PSFs matches.
-> This suggests the model is interpretable.
17. 1270 μm
8
0
0
μ
m
9
3
0
μ
m
z
y
x
Lateral plane: high resolution
Axial plane: low resolution
In our test case, anisotropy was severe.
*Mouse cortical regions imaged with confocalfluorescence microscopy
Lateral
Axial
Lateral
18. Our model blindly restores the degraded axial resolution.
*Mouse cortical regions imaged with confocalfluorescence microscopy
10
5
0
0.15
0.45
Position (μm)
Signal
intensity
20 μm
Lateral
Axial
G
21. The trained network also recovers the suppressed details.
*Mouse cortical regions imaged with confocalfluorescence microscopy
50 μm
10 μm
20 μm
Input Network 90°-rotated
22. The resolution enhancement was consistent across the entire sample space.
*Mouse cortical regions imaged with confocalfluorescence microscopy
z
x
y
z
x
y
Input
Network
10 μm
23. The enhancement translates to more detailed reconstruction of neuronal morphologies.
*Mouse cortical regions imaged with confocalfluorescence microscopy
0.05
0.64
Raw
0.05
0.64
Super-resolved
Traced neurites
Traced neurites
**Neurites traced with NeuroGPS-Tree andhuman correction.
** **
24. The method works for other biological samples - astrocytes.
*Rat cortical regions imaged with confocalfluorescence microscopy
20 μm
50 μm
XZ XZ
X
Z
YZ YZ
y
Z
XY
x
y
z
25. The method works for other biological samples – blood vessel.
*Ratcortical regions imaged with confocalfluorescence microscopy
Input Network
0
1
32. X
Y
Z
XY XZ YZ
In a poorly calibrated Open-top LSFM system, anisotropy is highly irregular across the sample space.
33. In a poorly calibrated Open-top LSFM system, anisotropy is highly irregular across the sample space.
Lateral image (shown as a reference)
Axial under-sampling + sample vibration from stage drift
Doubling effect by asynchronization between sweeps of excitation
source and detection path.
34. Our method addresses these artifacts which are not simply corrected by PSF-deconvolution.
x
z
50 μm
10 μm
35. The artifact correction in 3D provides better signal-to-background contrast.
y
z
x
y
z
x
XZ XZ
0.18
0.02
YZ
YZ
Raw Super-resolved
25μm
Editor's Notes
\
Intuition
DLG stands for deep linear generator.
Simulaiton involves 10,000 tubular objects.
For subsampling variations, the axial sub-sampling method of taking i-th slice takes place after the Gaussian blurring with standard deviation of 4.
FHWM mismatches were calculated from 317 tubular objects.
The dotted blue line is the axial input. The solid blue line is the axial output.
[CLICK TO SHOW THE PSNR GRAPH] The PSNR metrics were calculated for 31 ROIs of 140x140 microm, where the fluorescence distribution was deteced similarly between input and the rotated image. The network introduced a mean PSNR improvement of 2.42 dB per pair of input ROI versus output ROI.
The neurites were traced using a state-of-the-art neuronal tracing method for instance segmentation, NeuroGPS-tree, followed by human correction. The tracing was done blindly without knowledge of their image counterparts.
\
The real benefit of the light sheet microscopy when it comes to testing with deconvolutional capability is that you can start from almost near-isotropy and adjust the degree of anisotropy by altering the imaging conditions, such as the thickness of the light-sheet.
First, by taking care of the fine resolution capability of Open-top light sheet fluorescence microscopy, set the imaging condition to be isotropic as closely as possible. Then blur only the axial plane of the image volume by convolving with a Gaussian blurring in the z-axis. The gaussian blurring is set with the standard deviation of 10.
As we already know the PSF model of this image degradation, we can also compare the results to the ground-truth (before the image degradation) and also a classical deconovlution algorithm such as Richard-Lucy deconvolution algorithm, based on this PSF. What’s interesting is, while RL-deconvolved definitely show the effect of deconvolution, it also fails to reconstruct the fine details, which are mostly captured in the high frequency information that gets low-pass filtered out when we apply the blurring in the first place. In comparison, our deep-learning-based method excels at recontructing this lost high frequency information based on the local cues, as esepcially well demonstrated in the second column of the ROIs.
PSNR, MS-SSIM, and BRISQUE metrics are calculated for 8 XZ-plane MIP images that are in depths of 17.5 μm. BRISQUE is a non-reference-aware quantification to measure how naturally perceptible the image is. A lower score means more naturally perceptible.
The real benefit of the light sheet microscopy when it comes to testing with deconvolutional capability is that you can start from almost near-isotropy and adjust the degree of anisotropy by altering the imaging conditions, such as the thickness of the light-sheet.
Imaging a large-scale sample at a fine resolution, such as imaging a whole mouse brain at a sub-micrometer resolution, may introduce an aggregate of unexpected image artifacts that are not noticeable in coarser resolution. In LSM microscopy, these artifacts are often by-products of a microscope that is not properly calibrated or installed. In particular, standard OT-LSM systems require the excitation path and the imaging path to be perpendicular to each other and may introduce distortions to the image quality that are uneven between XZ plane and Y Z plane, although this anisotropy could be relaxed by tightly focused excitation.
Our system is now set up to include multiple imaging artifacts, which include not only the blurring artifacts by the spherical aberration that is caused by the refractive index mismatch between air and immersion medium, but other artifacts that span non-uniformly across the image space: for example, image doubling artifacts by a missed synchronization between the sweeping of the excitation laser and the rolling shutter of the detection sensor, or motion blur artifacts from physical sample drifts by the motorized stage.
ROIs show the examples of artifacts that are most likely caused the axial under-sampling and the motion blur artifacts from physical sample drifts by the motorized stage.