Noisy multimodal brain image registration using markov random field model

In this paper, a new scheme is proposed to register brain Magnetic Resonance (MR) and Computed Tomography (CT) images under a noisy environment. The scheme is developed based on the notion of mutual information and this is extended to the feature space-based registration. The 2D joint histogram of the noisy MR and CT images has been computed and this is assumed to be the degraded version of the true 2D joint histogram that corresponds to the CT and MR images. The true 2D joint histogram is modeled as Markov Random Field (MRF) model and is estimated from the degraded one by formulating the problem in Maximum A Posterior (MAP) estimation framework. The MAP estimates are obtained by the combination of Simulated Annealing (SA) and Iterated Conditioned Mode (ICM) algorithms. The Mutual Information (MI) at different values of the parameter vector. are computed from the estimated joint histogram and is maximized to determine the optimal registration parameter. The MI curve is modeled as multivariate Gaussian distribution and the registration parameter vector is obtained with a confidence interval. The scheme has been tested with the slices of patients from the Retrospective Image Registration Evaluation (RIRE) database. The efficacy of the scheme on MR and CT volumes is demonstrated by rendering the registered slices. The proposed algorithm is successfully tested with different degrees of Gaussian, Rician and Rayleigh noise conditions. Performance measures such as parameter estimation errors, error statistics and landmark based validation are used to demonstrate the efficacy of the proposed algorithm. The scheme is found to exhibit improved performance, especially at high noisy conditions as compared to the existing schemes.

Automated summary

This paper proposes a new scheme for registering brain MR and CT images under noisy conditions. The scheme is based on the concept of mutual information and extends it to feature space-based registration. The 2D joint histogram of the noisy MR and CT images is computed and assumed to be a degraded version of the true joint histogram. The true joint histogram is modeled as a Markov Random Field and estimated from the degraded version using MAP estimation. The optimal registration parameter is determined by maximizing the mutual information, which is modeled as a multivariate Gaussian distribution. The scheme is tested with patient slices from the RIRE database and demonstrates improved performance compared to existing schemes, especially in high noisy conditions.