A Comparative Evaluation of Unconstrained Optimization Methods Applied to the Thermal Tomography Problem

In cancer hyperthermia treatments, it is important to be able to predict complete tissue temperature fields from sampled temperatures taken at the limited number of locations allowed by clinical constraints. An initial attempt to do this automatically using unconstrained optimization techniques to minimize the differences between experimental temperatures and temperatures predicted from treatment simulations has been previously reported [1]. This paper reports on a comparative study which applies a range of different optimization techniques (relaxation, steepest descent, conjugate gradient, Gauss, Box-Kanemasu, and Modified Box-Kanemasu) to this problem. The results show that the Gauss method converges more rapidly than the others, and that it converges to the correct solution regardless of the initial guess for the unknown blood perfusion vector. A sensitivity study of the error space is also performed, and the relationships between the error space characteristics and the comparative speeds of the optimization techniques are discussed.