Public defence in Mathematics, M.Sc. (Tech) Juha-Pekka Puska

Title of the doctoral thesis: Bayesian Optimal Experimental Design in Imaging

Doctoral student: Juha-Pekka Puska
Opponent: Assist. Prof. Xun Huan, University of Michigan, USA
Custos: Prof. Nuutti Hyvönen, Aalto University School of Science, Department of Mathematics and Systems Analysis

Enabling optimal experimental design leads to more efficient use of data in imaging 

Optimal experimental design is the process of looking for experimental setups, which produce data that is as informative as possible, considering the purpose of the experiment. In the Bayesian setting this corresponds to an experimental design that produces probability distributions for the unknown variable, that contain the least amount of uncertainty. When finding an optimal design, it is required to compute the expected value of the utility of the experiment over the possible realizations of the measurement data and the unknown. 

The work in the thesis tackles the challenges that arise when applying optimal experimental design methods to imaging problems. Typical for imaging problems is the high dimensionality of the variables, and the large amount of data, which directly affect the computational demands. Computing the expectation over the experimental designs requires the evaluation of the high-dimensional integral, so the computational demands place restrictions on the dimension of the problem itself. 

Using suitable approximation methods, the computational demands can be lowered, and thus new applications of optimal experimental design can be opened. This work has concentrated especially on linear and sequential problems, for which a model example is computational tomography. A special challenge is the situation where the unknown variable is not assumed to be normally distributed, since in this situation the expected utility for the experiment does not have a form that can be easily evaluated. 

The work presents algorithms for efficient computations of optimal designs and applies these algorithms to medical imaging in simulated studies. The simulations show that optimized designs produce better reconstructions compared to reference setups, and thus optimal experimental design methods can produce better results in many imaging applications.

Thesis available for public display 10 days prior to the defence at: 
https://aaltodoc.aalto.fi/doc_public/eonly/riiputus/

Contact details:

Doctoral theses in the School of Science: https://aaltodoc.aalto.fi/handle/123456789/52