Damon Hyde

Location: ECE Conference Room, 442 Dana Research Building Abstract:

Statistical Models and Structured Regularization for Fluorescence Molecular Tomography

Date: Monday, Sept 29th
ECE Conference Room

Abstract

Fluorescence molecular tomography (FMT) is an optical imaging technique that uses near infrared light to localize and quantify textit{in vivo} distributions of fluorescent probes targeting biochemical markers such and genes, proteins, and enzymes. In this thesis, we examine three aspects of the FMT reconstruction problem: statistical data modeling in the context of normalized fluorescence imaging, methods for the use of prior structural information arising from multi-modal FMT-CT imaging, and techniques to compensate for errors in that prior information. We derive a probabilistic model for normalized fluorescence data and use this model as the basis for reconstruction. This eliminates errors and human biases introduced by manual data thresholding and is shown to yield improved reconstructions with greater consistency. To improve upon the resolution limits of stand-alone FMT, we examine modeling and regularization that incorporates structural prior information available from data acquired by a complementary imaging modality such as CT or MRI. We show that improved diffusion models using average tissue optical properties can result in subsequently improved reconstructions. A two step inversion approach is then presented, using the solution to an anatomically defined low dimensional problem as the basis of a spatially varying regularization term for the full resolution problem. Results are presented for both simulated and textit{in vivo} data, in the context of imaging a mouse model of Alzheimer's disease. Such diffuse targets are difficult to reconstruct with standalone FMT, thus highlighting the utility of the multimodal approach. Results are correlated with post mortem fluorescence measurements, and show a high degree of correlation between reconstruction intensity and observed fluorescence. Finally, two methods are presented to address situations where the prior information and underlying fluorescence share similar, but not identical, structure. The first uses differential equations to derive a Gaussian prior model for the fluorescence image. The incorporation of boundary conditions between physical regions allows information to cross anatomic boundaries, and helps compensate for boundary misplacement. The second approach uses the sparsity inducing properties of 1-norm minimization to localize the boundary within an uncertainty region around its initial position. Both approaches are tested using a range of 2-D simulated experiments.

Thesis Committee:
Prof Dana Brooks (co-advisor)
Prof Eric Miller, Tufts (co-advisor)
Prof Mark Niedre
Prof Vasilis Ntziachristos, Tech. Univ. Munich and NU (co-advisor)