A Comparative Study of Reinforcement Learning Algorithms Applied to Medical Image Registration

Nowadays, several medical procedures depend on the comparison and combination of images obtained in different modalities (magnetic resonance, computed tomography, PET, among others). Image registration is a geometric transformation process to align two or more images. It is necessary to have robust algorithms to find the best parameters of transformation in order to achieve accurate registrations. Reinforcement learning allows to train an agent through direct environment interaction, to achieve a goal. In this work, a comparison of the performance of Q-learning and Deep-Q with its variants is presented. Brain magnetic resonance images are used in 2D domain considering rigid deformations. The comparison is based on the reward values, computing the Pearson correlation factor in monomodal registration and Mutual information in multimodal registration, obtained during the learning process. It is also considered an error measure between the target parameters and the achieved ones. Finally, a backup memory criterion is proposed to train the Q-Network methods. Experimental results show a successfully behavior in all cases, but performance is improved when the proposed criterion is applied.