Search in sources :

Example 6 with FisherInformationMatrix

use of uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix in project GDSC-SMLM by aherbert.

the class PoissonGradientProcedureTest method crlbIsHigherWithPrecomputed.

@SeededTest
void crlbIsHigherWithPrecomputed(RandomSeed seed) {
    final int iter = 10;
    final UniformRandomProvider rng = RngUtils.create(seed.getSeed());
    final ErfGaussian2DFunction func = (ErfGaussian2DFunction) GaussianFunctionFactory.create2D(1, 10, 10, GaussianFunctionFactory.FIT_ERF_FREE_CIRCLE, null);
    final double[] a = new double[1 + Gaussian2DFunction.PARAMETERS_PER_PEAK];
    final int n = func.getNumberOfGradients();
    // Get a background
    final double[] b = new double[func.size()];
    for (int i = 0; i < b.length; i++) {
        b[i] = nextUniform(rng, 1, 2);
    }
    for (int i = 0; i < iter; i++) {
        a[Gaussian2DFunction.BACKGROUND] = nextUniform(rng, 0.1, 0.3);
        a[Gaussian2DFunction.SIGNAL] = nextUniform(rng, 100, 300);
        a[Gaussian2DFunction.X_POSITION] = nextUniform(rng, 4, 6);
        a[Gaussian2DFunction.Y_POSITION] = nextUniform(rng, 4, 6);
        a[Gaussian2DFunction.X_SD] = nextUniform(rng, 1, 1.3);
        a[Gaussian2DFunction.Y_SD] = nextUniform(rng, 1, 1.3);
        final PoissonGradientProcedure p1 = PoissonGradientProcedureUtils.create(func);
        p1.computeFisherInformation(a);
        final PoissonGradientProcedure p2 = PoissonGradientProcedureUtils.create(OffsetGradient1Function.wrapGradient1Function(func, b));
        p2.computeFisherInformation(a);
        final FisherInformationMatrix m1 = new FisherInformationMatrix(p1.getLinear(), n);
        final FisherInformationMatrix m2 = new FisherInformationMatrix(p2.getLinear(), n);
        final double[] crlb1 = m1.crlb();
        final double[] crlb2 = m2.crlb();
        Assertions.assertNotNull(crlb1);
        Assertions.assertNotNull(crlb2);
        // Arrays.toString(crlb2));
        for (int j = 0; j < n; j++) {
            Assertions.assertTrue(crlb1[j] < crlb2[j]);
        }
    }
}
Also used : ErfGaussian2DFunction(uk.ac.sussex.gdsc.smlm.function.gaussian.erf.ErfGaussian2DFunction) FisherInformationMatrix(uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix) UniformRandomProvider(org.apache.commons.rng.UniformRandomProvider) SeededTest(uk.ac.sussex.gdsc.test.junit5.SeededTest)

Example 7 with FisherInformationMatrix

use of uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix in project GDSC-SMLM by aherbert.

the class ApacheLvmFitter method computeFisherInformationMatrix.

@Override
protected FisherInformationMatrix computeFisherInformationMatrix(double[] y, double[] a) {
    final GradientCalculator c = GradientCalculatorUtils.newCalculator(function.getNumberOfGradients(), false);
    // Since we know the function is a Gaussian2DFunction from the constructor
    final double[][] I = c.fisherInformationMatrix(y.length, a, (NonLinearFunction) function);
    if (c.isNaNGradients()) {
        throw new FunctionSolverException(FitStatus.INVALID_GRADIENTS);
    }
    return new FisherInformationMatrix(I);
}
Also used : FisherInformationMatrix(uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix) GradientCalculator(uk.ac.sussex.gdsc.smlm.fitting.nonlinear.gradient.GradientCalculator)

Example 8 with FisherInformationMatrix

use of uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix in project GDSC-SMLM by aherbert.

the class FastMleSteppingFunctionSolver method computeFunctionFisherInformationMatrix.

@Override
protected FisherInformationMatrix computeFunctionFisherInformationMatrix(double[] y, double[] a) {
    // The fisher information is that for a Poisson process.
    // We must wrap the gradient function if weights are present.
    Gradient1Function f1 = (Gradient1Function) function;
    if (obsVariances != null) {
        f1 = OffsetGradient1Function.wrapGradient1Function(f1, obsVariances);
    }
    final PoissonGradientProcedure p = PoissonGradientProcedureUtils.create(f1);
    p.computeFisherInformation(a);
    if (p.isNaNGradients()) {
        throw new FunctionSolverException(FitStatus.INVALID_GRADIENTS);
    }
    return new FisherInformationMatrix(p.getLinear(), p.numberOfGradients);
}
Also used : OffsetGradient1Function(uk.ac.sussex.gdsc.smlm.function.OffsetGradient1Function) Gradient1Function(uk.ac.sussex.gdsc.smlm.function.Gradient1Function) PoissonGradientProcedure(uk.ac.sussex.gdsc.smlm.fitting.nonlinear.gradient.PoissonGradientProcedure) FisherInformationMatrix(uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix)

Example 9 with FisherInformationMatrix

use of uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix in project GDSC-SMLM by aherbert.

the class BaseFunctionSolver method computeDeviations.

@Override
public boolean computeDeviations(double[] y, double[] a, double[] parametersVariance) {
    // Use a dedicated solver optimised for inverting the matrix diagonal.
    final FisherInformationMatrix m = computeFisherInformationMatrix(y, a);
    setDeviations(parametersVariance, m);
    return true;
}
Also used : FisherInformationMatrix(uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix)

Example 10 with FisherInformationMatrix

use of uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix in project GDSC-SMLM by aherbert.

the class CreateData method reportAndSaveFittingLimits.

/**
 * Output the theoretical limits for fitting a Gaussian and store the benchmark settings.
 *
 * @param dist The distribution
 */
private void reportAndSaveFittingLimits(SpatialDistribution dist) {
    ImageJUtils.log(TITLE + " Benchmark");
    final double a = settings.getPixelPitch();
    final double[] xyz = dist.next().clone();
    final int size = settings.getSize();
    final double offset = size * 0.5;
    for (int i = 0; i < 2; i++) {
        xyz[i] += offset;
    }
    // Get the width for the z-depth by using the PSF Model
    final PsfModel psf = createPsfModel(xyz);
    psfModelCache = psf;
    double sd0;
    double sd1;
    if (psf instanceof GaussianPsfModel) {
        sd0 = ((GaussianPsfModel) psf).getS0(xyz[2]);
        sd1 = ((GaussianPsfModel) psf).getS1(xyz[2]);
    } else if (psf instanceof AiryPsfModel) {
        psf.create3D((double[]) null, size, size, 1, xyz[0], xyz[1], xyz[2], null);
        sd0 = ((AiryPsfModel) psf).getW0() * AiryPattern.FACTOR;
        sd1 = ((AiryPsfModel) psf).getW1() * AiryPattern.FACTOR;
    } else if (psf instanceof ImagePsfModel) {
        psf.create3D((double[]) null, size, size, 1, xyz[0], xyz[1], xyz[2], null);
        sd0 = ((ImagePsfModel) psf).getHwhm0() / Gaussian2DFunction.SD_TO_HWHM_FACTOR;
        sd1 = ((ImagePsfModel) psf).getHwhm1() / Gaussian2DFunction.SD_TO_HWHM_FACTOR;
    } else {
        throw new IllegalStateException("Unknown PSF: " + psf.getClass().getSimpleName());
    }
    final double sd = Gaussian2DPeakResultHelper.getStandardDeviation(sd0, sd1) * a;
    ImageJUtils.log("X = %s nm : %s px", MathUtils.rounded(xyz[0] * a), MathUtils.rounded(xyz[0], 6));
    ImageJUtils.log("Y = %s nm : %s px", MathUtils.rounded(xyz[1] * a), MathUtils.rounded(xyz[1], 6));
    ImageJUtils.log("Z = %s nm : %s px", MathUtils.rounded(xyz[2] * a), MathUtils.rounded(xyz[2], 6));
    ImageJUtils.log("Width (s) = %s nm : %s px", MathUtils.rounded(sd), MathUtils.rounded(sd / a));
    final double sa = PsfCalculator.squarePixelAdjustment(sd, a);
    ImageJUtils.log("Adjusted Width (sa) = %s nm : %s px", MathUtils.rounded(sa), MathUtils.rounded(sa / a));
    ImageJUtils.log("Signal (N) = %s - %s photons", MathUtils.rounded(settings.getPhotonsPerSecond()), MathUtils.rounded(settings.getPhotonsPerSecondMaximum()));
    boolean emCcd;
    double totalGain;
    final double qe = getQuantumEfficiency();
    final double noise = getNoiseEstimateInPhotoelectrons(qe);
    double readNoise;
    if (CalibrationProtosHelper.isCcdCameraType(settings.getCameraType())) {
        final CreateDataSettingsHelper helper = new CreateDataSettingsHelper(settings);
        emCcd = helper.isEmCcd;
        totalGain = helper.getTotalGainSafe();
        // Store read noise in ADUs
        readNoise = settings.getReadNoise() * ((settings.getCameraGain() > 0) ? settings.getCameraGain() : 1);
    } else if (settings.getCameraType() == CameraType.SCMOS) {
        // Assume sCMOS amplification is like a CCD for the precision computation.
        emCcd = false;
        // Not required for sCMOS
        totalGain = 0;
        readNoise = 0;
    } else {
        throw new IllegalStateException("Unknown camera type: " + settings.getCameraType());
    }
    // The precision calculation is dependent on the model. The classic Mortensen formula
    // is for a Gaussian Mask Estimator. Use other equation for MLE. The formula provided
    // for WLSE requires an offset to the background used to stabilise the fitting. This is
    // not implemented (i.e. we used an offset of zero) and in this case the WLSE precision
    // is the same as MLE with the caveat of numerical instability.
    // Apply QE directly to simulated photons to allow computation of bounds
    // with the captured photons...
    final double min = settings.getPhotonsPerSecond() * qe;
    final double max = settings.getPhotonsPerSecondMaximum() * qe;
    final double lowerP = Gaussian2DPeakResultHelper.getPrecision(a, sd, max, noise, emCcd);
    final double upperP = Gaussian2DPeakResultHelper.getPrecision(a, sd, min, noise, emCcd);
    final double lowerMlP = Gaussian2DPeakResultHelper.getMLPrecision(a, sd, max, noise, emCcd);
    final double upperMlP = Gaussian2DPeakResultHelper.getMLPrecision(a, sd, min, noise, emCcd);
    final double lowerN = getPrecisionN(a, sd, min, MathUtils.pow2(noise), emCcd);
    final double upperN = getPrecisionN(a, sd, max, MathUtils.pow2(noise), emCcd);
    if (settings.getCameraType() == CameraType.SCMOS) {
        ImageJUtils.log("sCMOS camera background estimate uses an average read noise");
    }
    ImageJUtils.log("Effective background noise = %s photo-electron " + "[includes read noise and background photons]", MathUtils.rounded(noise));
    ImageJUtils.log("Localisation precision (LSE): %s - %s nm : %s - %s px", MathUtils.rounded(lowerP), MathUtils.rounded(upperP), MathUtils.rounded(lowerP / a), MathUtils.rounded(upperP / a));
    ImageJUtils.log("Localisation precision (MLE): %s - %s nm : %s - %s px", MathUtils.rounded(lowerMlP), MathUtils.rounded(upperMlP), MathUtils.rounded(lowerMlP / a), MathUtils.rounded(upperMlP / a));
    ImageJUtils.log("Signal precision: %s - %s photo-electrons", MathUtils.rounded(lowerN), MathUtils.rounded(upperN));
    // Wrap to a function
    final PsfModelGradient1Function f = new PsfModelGradient1Function(psf, size, size);
    // Set parameters
    final double[] params = new double[5];
    // No background when computing the SNR
    // params[0] = settings.getBackground() * qe;
    params[1] = min;
    System.arraycopy(xyz, 0, params, 2, 3);
    // Compute SNR using mean signal at 50%. Assume the region covers the entire PSF.
    final double[] v = new StandardValueProcedure().getValues(f, params);
    final double u = FunctionHelper.getMeanValue(v, 0.5);
    final double u0 = MathUtils.max(v);
    // Store the benchmark settings when not using variable photons
    if (Double.compare(min, max) == 0) {
        ImageJUtils.log("50%% PSF SNR : %s : Peak SNR : %s", MathUtils.rounded(u / noise), MathUtils.rounded(u0 / noise));
        // Compute the true CRLB using the fisher information
        createLikelihoodFunction();
        // Compute Fisher information
        final UnivariateLikelihoodFisherInformationCalculator c = new UnivariateLikelihoodFisherInformationCalculator(f, fiFunction);
        // Get limits: Include background in the params
        params[0] = settings.getBackground() * qe;
        final FisherInformationMatrix m = c.compute(params);
        // Report and store the limits
        final double[] crlb = m.crlbSqrt();
        if (crlb != null) {
            ImageJUtils.log("Localisation precision (CRLB): B=%s, I=%s photons", MathUtils.rounded(crlb[0]), MathUtils.rounded(crlb[1]));
            ImageJUtils.log("Localisation precision (CRLB): X=%s, Y=%s, Z=%s nm : %s,%s,%s px", MathUtils.rounded(crlb[2] * a), MathUtils.rounded(crlb[3] * a), MathUtils.rounded(crlb[4] * a), MathUtils.rounded(crlb[2]), MathUtils.rounded(crlb[3]), MathUtils.rounded(crlb[4]));
        }
        benchmarkParameters = new BenchmarkParameters(settings.getParticles(), sd, a, settings.getPhotonsPerSecond(), xyz[0], xyz[1], xyz[2], settings.getBias(), totalGain, qe, readNoise, settings.getCameraType(), settings.getCameraModelName(), cameraModel.getBounds(), settings.getBackground(), noise, lowerN, lowerP, lowerMlP, createPsf(sd / a), crlb);
    } else {
        // SNR will just scale
        final double scale = max / min;
        ImageJUtils.log("50%% PSF SNR : %s - %s : Peak SNR : %s - %s", MathUtils.rounded(u / noise), MathUtils.rounded(scale * u / noise), MathUtils.rounded(u0 / noise), MathUtils.rounded(scale * u0 / noise));
        ImageJUtils.log("Warning: Benchmark settings are only stored in memory when the number of photons is " + "fixed. Min %s != Max %s", MathUtils.rounded(settings.getPhotonsPerSecond()), MathUtils.rounded(settings.getPhotonsPerSecondMaximum()));
    }
}
Also used : PsfModelGradient1Function(uk.ac.sussex.gdsc.smlm.model.PsfModelGradient1Function) UnivariateLikelihoodFisherInformationCalculator(uk.ac.sussex.gdsc.smlm.fitting.UnivariateLikelihoodFisherInformationCalculator) StandardValueProcedure(uk.ac.sussex.gdsc.smlm.function.StandardValueProcedure) FisherInformationMatrix(uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix) AiryPsfModel(uk.ac.sussex.gdsc.smlm.model.AiryPsfModel) ReadHint(uk.ac.sussex.gdsc.smlm.results.ImageSource.ReadHint) AiryPsfModel(uk.ac.sussex.gdsc.smlm.model.AiryPsfModel) GaussianPsfModel(uk.ac.sussex.gdsc.smlm.model.GaussianPsfModel) ImagePsfModel(uk.ac.sussex.gdsc.smlm.model.ImagePsfModel) PsfModel(uk.ac.sussex.gdsc.smlm.model.PsfModel) GaussianPsfModel(uk.ac.sussex.gdsc.smlm.model.GaussianPsfModel) CreateDataSettingsHelper(uk.ac.sussex.gdsc.smlm.ij.settings.CreateDataSettingsHelper) ImagePsfModel(uk.ac.sussex.gdsc.smlm.model.ImagePsfModel)

Aggregations

FisherInformationMatrix (uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix)16 PoissonGradientProcedure (uk.ac.sussex.gdsc.smlm.fitting.nonlinear.gradient.PoissonGradientProcedure)5 Gradient1Function (uk.ac.sussex.gdsc.smlm.function.Gradient1Function)3 OffsetGradient1Function (uk.ac.sussex.gdsc.smlm.function.OffsetGradient1Function)3 ConvergenceException (org.apache.commons.math3.exception.ConvergenceException)2 TooManyEvaluationsException (org.apache.commons.math3.exception.TooManyEvaluationsException)2 TooManyIterationsException (org.apache.commons.math3.exception.TooManyIterationsException)2 UniformRandomProvider (org.apache.commons.rng.UniformRandomProvider)2 Gradient2FunctionValueStore (uk.ac.sussex.gdsc.smlm.function.Gradient2FunctionValueStore)2 LikelihoodWrapper (uk.ac.sussex.gdsc.smlm.function.LikelihoodWrapper)2 PoissonGammaGaussianLikelihoodWrapper (uk.ac.sussex.gdsc.smlm.function.PoissonGammaGaussianLikelihoodWrapper)2 PoissonGaussianLikelihoodWrapper (uk.ac.sussex.gdsc.smlm.function.PoissonGaussianLikelihoodWrapper)2 PoissonLikelihoodWrapper (uk.ac.sussex.gdsc.smlm.function.PoissonLikelihoodWrapper)2 LeastSquaresBuilder (org.apache.commons.math3.fitting.leastsquares.LeastSquaresBuilder)1 Optimum (org.apache.commons.math3.fitting.leastsquares.LeastSquaresOptimizer.Optimum)1 LeastSquaresProblem (org.apache.commons.math3.fitting.leastsquares.LeastSquaresProblem)1 LevenbergMarquardtOptimizer (org.apache.commons.math3.fitting.leastsquares.LevenbergMarquardtOptimizer)1 ValueAndJacobianFunction (org.apache.commons.math3.fitting.leastsquares.ValueAndJacobianFunction)1 Array2DRowRealMatrix (org.apache.commons.math3.linear.Array2DRowRealMatrix)1 ArrayRealVector (org.apache.commons.math3.linear.ArrayRealVector)1