nick beavers project manager deconvolution from andy molnar software engineer
TRANSCRIPT
Nick BeaversProject Manager
Deconvolution
from
Andy MolnarSoftware Engineer
Basics of Fluorescent Imaging
What is Fluorescence Imaging?
• Unlike reflection or absorption imaging.
• Use molecules called fluorophores:
– Upon illumination by light of a specific wavelength they emit light of longer wavelengths.
Absorption (Excitation) Spectrum
Emission Spectrum
Wavelength
Stokes Shift
Lens
Light Source
Digital Camera
Emission Filter
Excitation Filter
Dichromatic Mirror
Objective Lens
Fluorescence Microscopy
Fluorescence Microscopy
Microscope Slide
Specimen
Objective Lens
Coverslip
Immersion Medium
Slide
Fluorescence Microscopy
Microscope Slide
Specimen
Coverslip
SlideThe Object: A collection of fluorophores small
enough to be considered point sources of light.
Image Formation
The Object (reality)
The Image (reality observed)
XY
XZ
XY
XZ
What Happened?
Aperture Diffraction
One Dimension - Diffraction
Aperature d
Point Source of Light
Two Dimensions - Airy Disk
PSF (Point Spread Function)
• Point in object Point Spread Function (PSF) in image.– The image is built out of PSFs, not points.
• Effect of PSF– Features blur together– Measurements are corrupted
• Deconvolution– Undo effect of PSF
Resolution and Blur
Effect of the PSF
• Object -> Image
PSF in 3 Dimensions
• In 3d Widefield, light from out of focus planes is detected in focal planes.
• In Confocal, pinholes help but there is still some z-blur especially when SA is present.
XY
XZ
Widefield Confocal
XY
XZ
The Problem
How to make a measured image better represent the real object ?
1) Counteract the PSF – that is put the light that spread (Airy disk) out back to its 2D location
2) Put light into the proper Z plane 3) Reduce the noise
This is best accomplished using deconvolution
Deconvolution ofFluorescence Images
Why Deconvolution?
• Increased Resolution
• Better Contrast
• Improved Signal-to-Noise Ratio
Noise
• Random nature of Photon Emission– Poission Noise– Usually dominant– Low light levels
• Digital Imaging System– Gaussian Noise
Noise
Widefield Example
• Max Projections
Confocal Example
Deconvolution Methods
DECON METHOD
FEATURE
Speed Subtractive Quantitative Requires PSF
Produces PSF S/N Noise
No/Nearest Neighbor Fast Yes No No No Lower Higher
Inverse Filter Moderate No No No No Higher Higher
Non-BlindConstrainedIterative *
Mod-Long No Yes Yes/No No Higher Lower
BlindConstrained Iterative *
Long No Yes No Yes Higher Lowest
* Can use measured or theoretical PSF
Imaging Equation
• Mathematical operation called a convolution.• The Microscope is a convolution operator.• The inverse of a convolution is deconvolution
=
Object convolved with the PSF equals the Image
ImagePSFObject
Iterative Non-Blind Deconvolution
Object Estimate ImagePSF Image Estimate
ErrorUpdate
-
First Iteration
(Fixed PSF)
Blind Deconvolution
Object Estimate ImagePSF Image Estimate
ErrorUpdate
-
First Stage - Image
(Adaptive PSF)
Blind Deconvolution
Object Estimate ImagePSF Image Estimate
ErrorUpdate
-
Second Stage - PSF
(Adaptive PSF)
• Benefits– Can adapt to imperfections in the Microscope.– Double optimization: Object and PSF are
modified– Can adapt to changes in refractive index,
specimen, environment.– Handles SA well. – No need to measure the PSF.– The only certain benefit of non-blind over
blind is that it is faster.
Benefits of Blind Deconvolution(Adaptive PSF)
Comparison
Original Image Nearest Neighbors
Inverse Filter Iterative
2D Widefield Example
Example: Slice View
slice 29
slice 37
raw deblurred
The Point Spread Function (PSF)
Determining the PSF
• Need a input PSF for • Inverse filter• Constrained Iterative• Statistical.
• Need PSF “first guess” for• Blind
PSF - Widefield vs. Confocal
XY
XZ
Widefield Confocal
XY
XZ
What effects the shape of PSF??
• Lens parameters• Modality• Specimen Parameters
XY
XZ
Widefield
XY
XZ
Confocal
Spherical Aberration
Rays at the periphery of the lens have a closer focal point
Rays entering near the center of the lens have a
farther focal point
Spherical Aberration
Z
n1 n1n2
Index of Refraction Mismatch
• Caused by– Changes in refractive
index– Thick specimens– Incorrect cover slip
thickness– Worse deeper in the
sample
• Solve by– Match RI– SA correction collar– Use biased PSF and
deconvolution
Spherical Aberration
PSF
• Theoretical• Measured• Blind
Theoretical
• Theoretical– Scope Properties
• NA, Lens RI, Pixel Spacing, Emissive Wavelength,
– Specimen properties• Specimen RI, Distance from Coverslip
Measured
• Bead Image• Must be taken under the same conditions as the
real sample will be.– Same embedding RI– Same distance from coverslip
Bead Image
Blind PSF
• PSF also changes as part of optimization• Blind: 2 step process in Autoquant
– Spherical Aberration Detection– PSF modified further as part of Deconvolution
• Determine PSF without measuring it.
PSF Comparison
Measured Theoretical Blind
Example
• What happens if you use the wrong PSF??
Raw Imagemax = 2,614
XY
XZ
Deconvolved no SAmax = 12,134
XY
XZ
Deconvolution with SA Correctionmax = 23,938
XY
XZ
Geometric Distortion
• With SA PSF changes with depth• Actual focus location also changes (PSF
Asymmetry)• Images look stretched.• Scale in Z• Many papers on this (Model 2010)
Deconvolution of Confocal Images
• Yes Confocal images can be deconvolved• Confocal images
– some blur.– noisy
• Accurate statistical model is important• Poisson model• Deconvolution reduces both blur and noise• Improves the Signal to Noise Ratio.• Multi-Photon, Spinning Disc, Structured
Illumination.
Confocal Examples
Collecting Optical Sections
• Setting the top and bottom of the scan
• Use Nyquist sampling– Sampling rate needed to collect all information
• Don’t break rules for convenience only out of necessity
Imaging Approach
• First try Widefield + deconvolution– Better for low light levels– Living cells– Proteins
• Next Confocal + deconvolution– Thick specimens (Tissues)
• Multiphoton + deconvolution– Very thick specimens
Deconvolution Tips
• Check the image stack’s metadata before decon.
• Use the line profile to prove that decon works.
• Use statistics to show…– Average image intensity remain relatively the same– Max pixel intensity can increase by 100 times.
Deconvolution Tips
• 99% of the time Blind is best.
• Severe spherical aberration correction is best done by CGC on the objective.
• Use an AOI for SAC testing or decon testing.
• AutoQuant 30 Day trial program:– Use it to your advantage!– Customers receive the entire product to evaluate.
Q&A
Questions?
Nick BeaversProject Manager
Deconvolution
from
Andy MolnarSoftware Engineer