Maximum-lesion-detectability reconstruction using penalized maximum likelihood

Statistical reconstruction methods based on the penalized maximum-likelihood (or maximum a posteriori) principle have been developed to improve the image quality in emission tomography. However, the penalty functions (or image priors) are often heuristically designed to encourage smoothness in reconstructed images. This paper presents a method to design the penalty function to improve lesion detection. We factorize the penalty function into pairwise potentials. Using the theoretical expressions that we have developed for evaluating lesion detectability, we are able to find the optimum weights of the pairwise potentials to achieve the maximum lesion detectability based on the numerical observers that are correlated to human performance.