Sensitivity to prior knowledge in optical tomographic reconstruction

The performance of reconstruction algorithms for near-IR optical tomography depends critically on the accuracy of the forward model used to evaluate the closeness of a given solution to that most consistent with the data (Maximum A-Posteriori criterion). Statistical photon noise can be accounted for theoretically, but there are also problems with inaccurate geometry, refractive index mismatching, and boundary effects. Sensitivity to such effects depends extensively on what measures are being used (time-varying intensity, integrated intensity, mean time, etc.). Although reconstructions have been obtained with a variety of data measures, including noise, they are often derived under strict assumptions about the accuracy of the model. In this paper we discuss the robustness of data measures and image reconstruction in the presence of model inaccuracies. In particular we consider robustness with respect to geometric errors in the modeling of the support of the solution and to the initial estimates for the starting solution vector.