Assignment 4 part 2: Measure resolution

From Course Wiki
Revision as of 16:24, 30 September 2017 by Steven Wasserman (Talk | contribs)

Jump to: navigation, search
20.309: Biological Instrumentation and Measurement

ImageBar 774.jpg

This is part 2 of Assignment 4.

Overview

Synthetic image of tiny microspheres used for measuring resolution.

One of the most commonly used definitions of the resolution limit $ R $ of an optical system is the distance between two point sources in the sample plane such that the peak of one source’s image falls on the first zero of the other source’s image. This particular definition is called the Rayleigh resolution.

The theoretical value of $ R $ is given by the formula

$ R=\frac{0.61 \lambda}{ \text{NA}} $,

where $ \lambda $ is the wavelength of light that forms the image, and NA is the numerical aperture of the optical system. The definition suggests a procedure for measuring resolution: make an image of a point source, measure the peak-to-trough distance in the image plane, and divide by the magnification. In this part of the lab, you will use a procedure inspired by this simple idea to estimate the resolution of your microscope. Instead of measuring the spot sizes with a ruler, you will use nonlinear regression to find best-fit parameters of a two dimensional Gaussian function that best approximates the digital images of (near) point sources that you will make. You will use the best-fit parameters from the regression to compute the resolution measurement.

One practical problem with this method is that true point sources are difficult to come by. If you were an astronomer testing a telescope, stars are readily available in the night sky, and they are very good approximations of point sources. Since there is no natural microscopic sample that is equivalent to the night sky, microscopists have to prepare a synthetic sample suitable for measuring resolution. Prehaps the most common method is to use a microscope slide sprinkled with tiny, fluorescent beads that have diameters in the range of 100-190 nm. These beads are small enough to be considered point sources. Unfortunately, beads small enough for this purpose are not very bright. Imaging them can be challenging. Your microscope must be very well aligned to get good results. (The images you make for this part of the lab will probably remind you of telescope images. If they don't, have an instructor take a look at your setup.)

Why fit a Gaussian instead of a Bessel function? Gaussians are more amenable to nonlinear regression because they are smoother and faster to evaluate than Bessel functions. In addition, the Gaussian is a very good approximation to the central bump of a Bessel function. It is straightforward to convert the Gaussian parameters to Rayleigh resolution. See Converting Gaussian fit to Rayleigh resolution for a discussion of the conversion.

In outline, the sequence of steps for the resolution measurement is:

  1. get a (real or synthetic) image of point sources.
  2. find the pixels that correspond to each microsphere and associate them into connected regions
  3. compute useful properties of the connected regions
  4. eliminate regions that are likely not image of a single microscphere
  5. use nonlinear regression to fit a Gaussian model function to the image of each microsphere
  6. compute summary statistics
  7. convert Gaussian parameter to Rayleigh resolution

We're already done with the first step. Use the image of 180 nm microsphere you generated in part 1 of this assignment.

Finding and computing properties of connected regions

The next step is to identify pixels of interest. Fortunately, this is a very simple matter because we went to great lengths to make an image that has high contrast, little background, and high SNR. The interesting pixels are the bright ones. A simple, global threshold works well — all the pixels brighter than a certain threshold are (probably) interesting. You might want to think a bit about how to choose the best threshold value. One complicating factor is that the illumination is usually brighter in the center of the image than at the edges, so a threshold that works well in the middle of the image may not work well toward the edges. (Protip: don't flat-field correct your PSF image.) The MATLAB Image Processing Toolbox includes a global threshold function. im2bw( I, level ) applies a global threshold to image I and returns a binary image (also called a bilevel image). The binary image is a matrix the same size as I that contains only ones and zeros. You guessed it — there are ones in locations where the pixel value was greater than level and zeroes everywhere else.

The function regionprops operates on binary images. It can identify connected regions and compute properties of those regions. The function FindBrightObjectsInImage below uses im2bw and regionprops to segment an image and compute properties of each connected region of pixels. Region properties to compute (e.g. area, eccentricity) are specified in a cell array of strings. Here is a complete list of the properties regionprops can compute.

function RegionProperties = FindBrightObjectsInImage( ...
        InputImage, GlobalThreshold, DilationRadius, PropertyList )

    mask = im2bw(InputImage, GlobalThreshold);
    mask = imclearborder(mask);               % eliminates connected regions that touch the edge
    mask = imdilate(mask, strel( 'disk', DilationRadius ));  % open up the mask a little
    
    RegionProperties = regionprops( mask, InputImage, PropertyList );

end

FindBrightObjectsInImage returns is a struct array with all of the computed properties —, an array where each element is a structure. The structure will contain one field for each of the properties in the PropertyList argument. The fields have the same name as the property. For example, if you included 'Centroid' in the property list and there were ten objects in the image, RegionProperties(3).Centroid would return a 1 x 2 matrix with the y and x coordinates of the centroid. If you wanted to create an N x 2 of all of the centroids, you can use MATLAB's : indexing syntax and a concatenation function: AllCentroids = vertcat( RegionProperties(:).Centroid );.

Try running FindBrightObjectsInImage on your synthetic image and examine the results. Put a breakpoint on the first line of FindBrightObjectsInImage. Use imshow to see how the mask is affected by imclearborder and imdilate. What is a good value for DilationRadius?

Using nonlinear regression to measure resolution

The task of measuring resolution would be super-simple if regionprops had a built-in method for measuring resolution. Unfortunately, it doesn't have one so we will have to write our own. We can use nonlinear regression to compute parameters for each of the PSF bead images.

Regression concepts review

Nonlinear regression is a method for finding the relationship between a dependent quantity $ O_n $ and an one or more independent variables, in this case the spatial coordinates of the image $ x_n $ and $ y_n $. The variables are related by a model function $ f(\beta, x_n, y_n) $. $ \beta $ is a vector of model parameters. The dependent variable $ O_n $ is measured in the presence of random noise, which is represented mathematically by a random variable $ \epsilon_n $. In equation form:

$ O_n=f(\beta, x_n, y_n)+\epsilon_n $.

The goal of regression is to determine a set of best-fit model parameters $ \hat{\beta} $ that match an observed data as closely as possible. Because the dependent variable includes noise, $ \beta $ cannot be determined exactly from the data. Increasing the number of observations or decreasing the magnitude of the noise tends to produce a more reliable estimate of $ \beta $.

Linear and nonlinear regression are similar in some aspects, but the two techniques have a fundamental difference. Nonlinear regression cannot be reduced to a single, deterministic formula like linear regression. Finding the optimal solution to a nonlinear regression is an iterative process. Nonlinear regression begins with an initial guess. Each iteration produces a more refined estimate of $ \beta $. The process stops when no better estimate can be found (or when something bad happens ... such as the solution not converging).

Ordinary nonlinear least squares regression assumes that:

  • the independent variables are known exactly, with zero noise,
  • the error values are independent and identically distributed,
  • the distribution of the error terms has a mean value of zero,
  • the independent variable covers a range adequate to define all the model parameters, and
  • the model function exactly relates $ O $ to $ x $ and $ y $.

These assumptions are almost never perfectly met in practice. It is important to consider how badly the regression assumptions have been violated when assessing the results of a regression.

Get ready …

As shown in the diagram below, you need four things to run a regression:

  1. a matrix containing the values of the independent variable(s);
  2. a vector containing the corresponding observed values of the dependent variable;
  3. a model function; and
  4. a vector of initial guesses for the model parameters.
Block diagram of nonlinear regression.

1 and 2: Independent variable and observations

In the regression we are about to do, the related quantities are x and y, the coordinates of each pixel in the image of a microsphere, and the corresponding pixel values (intensities) of each pixel. One important consideration is that the regression algorithm assumes that the independent variable has no (or very little) noise. Happily, the pixel coordinates are known with essentially zero error (the layout of pixels on the camera is extremely accurate), which means our regression will not violate this assumption. The pixel values, however, are subject to imaging noise. This makes it easy to determine the x and y are a good choice to use as independent variables and the pixel value will be the dependent variable.

The regionprops function can compute two properties, PixelList and PixelValues that will be useful for the regression. PixelList is an N x 2 matrix that contains the y and x coordinates of each pixel in the region. PixelValues is a vector of all the corresponding pixel values (in the same order).

Model function

nlinfit requires that the regression model be expressed as a function that takes two arguments and returns a single vector of predicted values. The model function must must have the form:

[ PredictedValues ] = ModelFunction( Beta, X )

The first argument, Beta, is a vector of model parameters. The second argument, X is a vector of independent variable values. The return value, PredictedValues, must must have the same size as X.

The MATLAB function Gaussian2DFitFunction defined below computes the two dimensional function that we will use to model the image of a PSF bead. Parameters is a 1x5 vector that contains the model parameters in this order: X center, Y center, amplitude, sigma, and offset.

function out = Gaussian2DFitFunction( Parameters, Coordinates )
    yCenter = Parameters(1);
    xCenter = Parameters(2);
    amplitude = Parameters(3);
    sigma = Parameters(4);
    offset = Parameters(5);
    
    out = amplitude * ...
        exp( -(( Coordinates(:, 1) - yCenter ).^2 + ( Coordinates(:, 2) - xCenter ).^2 ) ...
        ./ (2 * sigma .^ 2 )) + offset;
    
end

It's a good idea to test the model function out before you use it. The plot below shows four sets of curves generated by Gaussian2DFitFunction with different parameters. It's comforting to see that the curves have the expected shape.

Two dimensional Gaussians generated by Gaussian2DFitFunction.png

Initial guesses

nlinfit requires an initial value for each of the three model parameters, contained in a 1x5 vector. (nlinfit infers the number of model parameters from the size of the Initial guess vector.)

    pixelCountAboveHalf = sum( Values > ( ( min( Values ) + max( Values ) ) / 2 ) );
    sigmaInitialGuess = 0.8 * sqrt( pixelCountAboveHalf / 2 / pi / log(2) );

    initialGuesses = [ ...
        mean( Coordinates(:, 1) ), ... % yCenter
        mean( Coordinates(:, 2) ), ... % xCenter
        range( Values ), ... % amplitude
        sigmaInitialGuess, ... % sigma
        min( Values ) ]; % offset

Get set …

The first step of all regressions is to plot the observations and the model function evaluated with the initial guesses versus the independent variable on a single set of axes. Don't attempt to run nlinfit until you've done this plot. It is much easier to ensure that the arguments to nlinfit are plausible before you invoke it than to debug a screen full of cryptic, red text afterwards. Side effects of premature regression include confusion, waste of time, fatigue, irritability, alopecia, and feelings of frustration. Contact your professor if your regression lasts more than four hours. There is no chance that nlinfit will succeed if there is a problem with one of its arguments.

Go ahead ... do the plot.

plot3( Coordinates(:,1), Coordinates(:,2), Gaussian2DFitFunction( initialGuesses, Coordinates ), 'x' )
hold on
plot3( Coordinates(:,1), Coordinates(:,2), Values, 'x' )


2D Gaussian pre-regression plot.png

It looks like the initial guesses are good. Now we can proceed with confidence that the arguments to nlinfit are credible.

Go...

Here is a function that puts the whole nonlinear regression process together:

function BestFitParameters = Fit2dGaussian( Values, Coordinates )
    pixelCountAboveHalf = sum( Values > ( ( min( Values ) + max( Values ) ) / 2 ) );
    sigmaInitialGuess = 0.8 * sqrt( pixelCountAboveHalf / 2 / pi / log(2) );

    initialGuesses = [ ...
        mean( Coordinates(:, 1) ), ... % yCenter
        mean( Coordinates(:, 2) ), ... % xCenter
        range( Values ), ... % amplitude
        sigmaInitialGuess, ... % sigma
        min( Values ) ]; % offset

    BestFitParameters = nlinfit( Coordinates, Values, @Gaussian2DFitFunction, initialGuesses );
end

function out = Gaussian2DFitFunction( Parameters, Coordinates )
    yCenter = Parameters(1);
    xCenter = Parameters(2);
    amplitude = Parameters(3);
    sigma = Parameters(4);
    offset = Parameters(5);
    
    out = amplitude * ...
        exp( -(( Coordinates(:, 1) - yCenter ).^2 + ( Coordinates(:, 2) - xCenter ).^2 ) ...
        ./ (2 * sigma .^ 2 )) + offset;
    
end

Loop through all the connected regions

function [ Resolution, StandardError, BestFitData ] = MeasureResolutionFromPsfImage( ImageData )    
    % TODO list:
    % 1. think of a good way to pick the threshold
    % 2. figure out how to eliminate images that are not single beads
    
    objectProperties = FindBrightObjectsInImage( ImageData, 0.5, 2, { 'Centroid', 'PixelList', 'PixelValues' } );
        
    figure(1);
    imshow( ImageData );
    LabelObjectsInImage( objectProperties );
    
    % INSERT CODE TO ELIMINATE BAD OBJECTS HERE

    BestFitData = zeros( numel(objectProperties), 5);
    
    figure(2);
    
    % use nlinfit to fit a Gaussian to each object
    for ii = 1:length(objectProperties)
        
        BestFitData(ii, :) = Fit2dGaussian( objectProperties(ii).PixelValues, objectProperties(ii).PixelList );
        
        % plot data, initial guess, and fit for each peak
        figure(2)
        clf
        
        % generate a triangle mesh from the best fit solution found by 
        % nlinfit and plot it
        gd = delaunay( objectProperties(ii).PixelList(:,1), ...
            objectProperties(ii).PixelList(:,2) );
        trimesh( gd, objectProperties(ii).PixelList(:,1), ...
            objectProperties(ii).PixelList(:,2), ...
            Gaussian2DFitFunction(BestFitData(ii, :), ...
            objectProperties(ii).PixelList ) )
        hold on
        
        % plot image data
        plot3( objectProperties(ii).PixelList(:,1), ...
            objectProperties(ii).PixelList(:,2), ...
            objectProperties(ii).PixelValues, 'gx', 'LineWidth', 3)
        title(['Image data vs. Best Fit for Object Number ' num2str(ii)]);
        drawnow
    end
    
    Resolution = mean( BestFitData(:,4) ) ./ .336;
    StandardError = std( BestFitData(:,4) ./ .336 ) ./ sqrt( size( BestFitData, 1 ) );
end

function out = Gaussian2DFitFunction( Parameters, Coordinates )
    yCenter = Parameters(1);
    xCenter = Parameters(2);
    amplitude = Parameters(3);
    sigma = Parameters(4);
    offset = Parameters(5);
    
    out = amplitude * ...
        exp( -(( Coordinates(:, 1) - yCenter ).^2 + ( Coordinates(:, 2) - xCenter ).^2 ) ...
        ./ (2 * sigma .^ 2 )) + offset;
    
end

function LabelObjectsInImage( objectProperties )
    labelShift = -9;
    fontSize = 10;
    
    for ii = 1:length(objectProperties)
        unweightedCentroid = objectProperties(ii).Centroid;
        text(unweightedCentroid(1) + labelShift, unweightedCentroid(2), ...
            num2str(ii), 'FontSize', fontSize, 'HorizontalAlignment', ...
            'Right', 'Color', [0 1 0]);
    end

end

One more thing …

It's frequently the case that a few of the beads in a real PSF image are not good candidates for measuring resolution. For example, there are sometimes two beads that too close together to separate. Sometimes, there are also aggregates of multiple beads in the picture. Identify some useful properties that regionprops computes for sorting out the bad regions and write code to eliminate them.

A bit of MATLAB syntax that you might find useful is this: you can remove an element from an array by assigning it to be the empty value []. Some examples:

RegionProperties(3) = []; % removes the third element of the struct array and reduces its size by 1
RegionProperties( [ 0 0 0 1 0 1 0 0 1 0 ] ) = []; % removes the 4th, 6th, and 9th elements and reduces size by 3

Measuring resolution

EstimateResolutionFromPsfImage takes a point-source image and estimates the resolution of an optical system. It uses the built-in MATLAB function im2bw to locate bright regions and regionprops to measure attributes of each connected region of bright pixels. After rejecting outliers, the function uses nlinfit to estimate best fit Gaussian parameters for each bright spot. The optional second argument controls the rejection range for outliers.

There are four subfunctions that should be included in the same m-file as EstimateResolutionFromPsfImage.


function [ Resolution, StandardError, BestFitData ] = MeasureResolutionFromPsfImage( ImageData )

    
    % convert the image to souble precision, if needed
    if ~isa( ImageData, 'double' )
        ImageData = double( ImageData );
    end
    
    % TODO list:
    % 1. think of a good way to pick the threshold
    % 2. figure out how to eliminate images that are not single beads
    
    objectProperties = FindBrightObjectsInImage( ImageData, 0.5, 2, { 'Centroid', 'PixelList', 'PixelValues' } );
    
    labelShift = -9;
    fontSize = 10;
    
    figure
    imshow( ImageData );
    
    for ii = 1:length(objectProperties)
        unweightedCentroid = objectProperties(ii).Centroid;
        text(unweightedCentroid(1) + labelShift, unweightedCentroid(2), ...
            num2str(ii), 'FontSize', fontSize, 'HorizontalAlignment', ...
            'Right', 'Color', [0 1 0]);
    end
    
    % INSERT YOUR CODE TO ELIMINATE BAD OBJECTS HERE

    BestFitData = cell(1, length(objectProperties));
    
    figure;
    
    % use nlinfit to fit a Gaussian to each object
    for ii = 1:length(objectProperties)
        % initial guess for sigma based on area of bright spot
        maximumPixelValue = max( objectProperties(ii).PixelValues );
        darkPixelValue = median( objectProperties(ii).PixelValues );
        pixelCountAboveHalf = sum( objectProperties(ii).PixelValues > .5 *  ( maximumPixelValue + darkPixelValue ) );
        sigmaInitialGuess = 0.8 * sqrt( pixelCountAboveHalf / 2 / pi / log(2) );
        
        initialGuesses = [ ...
            objectProperties(ii).Centroid(1), ... % yCenter
            objectProperties(ii).Centroid(2), ... % xCenter
            max(objectProperties(ii).PixelValues) - min(objectProperties(ii).PixelValues), ... % amplitude
            sigmaInitialGuess, ... % (objectProperties(ii).BoundingBox(3) - 6) / 4, ... % sigma
            min(objectProperties(ii).PixelValues) ];
      
        BestFitData{ii} = nlinfit( objectProperties(ii).PixelList, objectProperties(ii).PixelValues, @Gaussian2DFitFunction, initialGuesses );
        
        % plot data, initial guess, and fit for each peak
        figure(2)
        clf
        
        % generate a triangle mesh from the best fit solution found by 
        % nlinfit and plot it
        gd = delaunay( objectProperties(ii).PixelList(:,1), ...
            objectProperties(ii).PixelList(:,2) );
        trimesh( gd, objectProperties(ii).PixelList(:,1), ...
            objectProperties(ii).PixelList(:,2), ...
            Gaussian2DFitFunction(BestFitData{ii}, ...
            objectProperties(ii).PixelList ) )
        hold on
        
        % plot initial guesses -- commented out to make plots less
        % cluttered. put this back in to debug initial guesses
        % plot3( objectProperties(ii).PixelList(:,1), ...
        %   objectProperties(ii).PixelList(:,2), ...
        %   Gaussian2DFitFunction(initialGuesses, ...
        %   objectProperties(ii).PixelList ), 'rx' )
        
        % plot image data
        plot3( objectProperties(ii).PixelList(:,1), ...
            objectProperties(ii).PixelList(:,2), ...
            objectProperties(ii).PixelValues, 'gx', 'LineWidth', 3)
        title(['Image data vs. Best Fit for Object Number ' num2str(ii)]);
    end
    
    allPeakData = vertcat( BestFitData{:} );
    if( ~isempty( allPeakData ) )
        Resolution = mean( allPeakData(:,4) ) ./ .336;
        StandardError = std( allPeakData(:,4) ./ .336 ) ./ sqrt( length( BestFitData ) );
    else
        Resolution = NaN;
        StandardError = NaN;
    end
end

function out = Gaussian2DFitFunction( Parameters, Coordinates )
    yCenter = Parameters(1);
    xCenter = Parameters(2);
    amplitude = Parameters(3);
    sigma = Parameters(4);
    offset = Parameters(5);
    
    out = amplitude * ...
        exp( -(( Coordinates(:, 1) - yCenter ).^2 + ( Coordinates(:, 2) - xCenter ).^2 ) ...
        ./ (2 * sigma .^ 2 )) + offset;
    
end

Testing the code

Example image processing on PSF beads to determine microscope resolution.
Example Gaussian fit of a PSF bead fluorescence emission profile to estimate microscope resolution.


Pencil.png

Use the synthetic image code you developed in part 1 of this assignment to test the EstimateResolutionFromPsfImage function using synthetic images of 170 nm fluorescent microspheres over a range of numerical apertures from 0.1 to 1.0. Plot the results, actual resolution versus measured resolution. Turn in your code and the plot.


Measure the resolution of your microscope

  1. Make an image of a sample of 170 nm fluorescent beads with the 40X objective. (Several dozens to hundreds of PSF spheres should be captured in your image.)
    • Use 12-bit mode on the camera and make sure to save the image in a format that preserves all 12 bits.
    • Ensure that the image is exposed properly.
      • Over-exposed images will give inaccurate results.
      • Under-exposed images will be difficult to process and yield noisy results.
    • This procedure is extremely sensitive to the focus adjustment.
    • To minimize photobleaching, do not expose of the beads to the light source and longer than necessary.
    • Be sure to save the image and the histogram for your lab report.
  2. Use image processing functions to locate non-overlapping, single beads in the image.
  3. Use nonlinear regression to fit a Gaussian to each bead image.
  4. Convert the Gaussian parameters to resolution.


Pencil.png

Report the resolution you measured and discuss sources of error in the measurement.


Back to 20.309 Main Page
Back to Assignment 4 Overview
Back to Assignment 4 Part 1
On to Assignment 4 Part 3