Dynamic Regularization for the Restoration of PET Images

摘要

Among the variety of methods so far proposed, the deconvolving process might be one of the most important techniques in improving medical imaging quality. In this paper, we will discuss the inverse problem of the deconvolving process within the framework of the network inversion technique, with respect to the neural network model for the partial volume effect arising in positron emission tomography(PET). To stabilize the inverse solution, we incorporate such a priori knowledge as the histological information given by MR images and the smoothness in the distribution of blood flow, which can be realized by Tikhonov regularization. In particular, in this paper we propose a "dynamic regularization technique" in which the regularizing parmeter is changed dynamically in each stage of the ongoing iterative optimization procedure. We explain the method and demonstrate the effectiveness of dynamic regularization with respect to the brain PET image restoration process, presenting successful results.