1 / 25

Numerical Modeling for Image Reconstruction

Numerical Modeling for Image Reconstruction. Subha Srinivasan 11/2/2009. Definition of Inverse Problem. Definition: Given a distribution of sources and a distribution of measurements at the boundary dΩ, finding the tissue parameter distribution within domain Ω. Expressed as x = F -1 (y).

Download Presentation

Numerical Modeling for Image Reconstruction

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Numerical Modeling for Image Reconstruction SubhaSrinivasan 11/2/2009

  2. Definition of Inverse Problem.. • Definition: Given a distribution of sources and a distribution of measurements at the boundary dΩ, finding the tissue parameter distribution within domain Ω. • Expressed as x = F-1(y)

  3. Ways of Solving Inverse Problems • Back-projection methods • Perturbation methods • Non-linear optimization methods

  4. Back-projection methods • Assumes that each projection provides a nearly independent measurement of the domain. • Assumes that light travels in a straight line: not true with tissue unless scattering is isolated fan beam detectors patient x-rays Filtered Back-projection method: [measurements] = [attenuation op.] [object ] [image] = [attenuation op.]T [filter] [measurements]

  5. Linear reconstruction for Change in Optical Properties • x = F-1(y) is a non-linear problem: can be linearized using Taylor’s series expansion if initial estimate is close to actual values: Jacobian matrix Reconstructing for changes rather than absolute values

  6. Structure of Jacobian Ф = I e-i(ωt+θ) I = signal amplitude θ = signal phase m = [ma, D] Absorption coeff. Diffusion Coeff.

  7. Shape of Jacobian Calculated by: 1)Perturbation Method 2)Direct Analytic Jacobian 3) Adjoint method

  8. Adjoint method for Jacobian Calculation Dehghani notes

  9. Solving • Linearizing change in intensity: born approximation • Linearizing change in log intensity: Rytov approximation • Inverting J: large, under-determined and ill-posed: some standard methods can be used • Truncated SVD, Tikhonov regularization, Algebriac reconstruction techniques (ART) & Conjugate Gradient methods are commonly used

  10. Terminology: Inverse Problem • Ill-posed–Small changes in the data can cause large changes in the parameters. • Ill-conditioned–The condition number (ratio of largest singular value to smallest singular value) is large, which implies the inverse solution would not be unique. • Ill-determined–(or under-determined) The number of independent equations are smaller than number of unknowns.

  11. Deriving Update Equation using Least Squares Minimization • Minimizing error functional: • Setting derivative to zero: • Taylor’s approximation • Rewriting: • Substituting: • Update equation:

  12. Assumptions of Levenberg-Marquardt Minimization: • JTJ is positive-definite • Initial guess must be close to actual solution • Update equation does not solve first-order conditions unless α = 0 *Yalavarthy et. al., Medical Physics, 2007

  13. Tikhonov Minimization: Key idea is to introduce apriori assumptions about size and smoothness of desired solution: L is dimensionless common choice: L = I (the identity matrix) *Tikhonov et. al, 1977; Tarantola SIAM 2004. *Yalavarthy et. al., Medical Physics, 2007

  14. Tikhonov Minimization Advantage: • parameters within the minimization scheme => stability Limitation: • it requires a prior opinion about the noise characteristics of the parameter and data spaces (for λ)

  15. Flow-chart for Iterative Image Reconstruction

  16. Choosing Regularization: L-curve criterion • Convenient graphical tool for displaying trade-off between size of solution and its fit to the given data as λ varies. • λ can also be chosen empirically or based on parameter/data values. Hansen, ‘L-curve and its use in numerical treatment of inverse problems’

  17. Reconstruction Results • Simulated Measurements, 5% Noise

  18. Recovery of Absorption

  19. Recovery of Scattering

  20. Spectral Image Reconstruction Data from Boulnois et al, Hale & Quarry, figure from thesis Srinivasan et al

  21. Spectral Image Reconstruction Relationships between Jsp & J can be obtained Details, refer to Srinivasan et al, AO, 2005

  22. Simulations show Reduced Cross-talk in spectral images HbT(μM) StO2(%) Water (%) Scatt Ampl. Scatt Power True • Data generated from a tumor-simulating phantom using FEM forward model, with 1% random-Gaussian noise added. Spectral Conv. • Spectral Method: Smoother Images; 15.3 % mean error compared to 43% (conv. Method). • Reduced Cross-talk between HbO2 and water: from30% (conv.) to 7% (spectral). • Accuracy in StO2 accurate (<1% error) Srinivasan et al, PhD thesis, 2005

  23. Results from Image Reconstruction:Experimental Data Brooksby, Srinivasan et al, Opt Lett, 2005

  24. References • Gibson et al, Phy Med Bio: 50 : 2005: A review paper • Paulsen et al, Med Phy: 22(6): 1995: first results from image-reconstruction in DOT • Yalavarthy et al, Med Phy: 34(6): 2007: good explanation of math • Brooksby et al, IEEE Journal of selected topics in quantum electronics: 9(2): 2003: good reference for spatial priors • Hansen: ‘Rank deficient and discrete ill-posed problems’: SIAM: 1998: good reference for tikhonov/l-curve • Srinivasan et al, Appl Optics: 44(10): 2005: reference for spectral priors • Press et al: ‘Numerical Recipes in Fortran 77’: II edition: 1992: great book for numerical folks!

More Related