Abstract

Rician noise and intensity inhomogeneity are two common types of image degradation that manifest in the acquisition of magnetic resonance imaging (MRI) system images of the brain. Many noise reduction and intensity inhomogeneity correction algorithms are based on strong parametric assumptions. These parametric assumptions are generic and do not account for salient features that are unique to specific classes and different levels of degradation in natural images.

This thesis proposes the 4-neighborhood clique system in a layer-structured Markov random field (MRF) model for noise estimation and noise reduction. When the test image is the only physical system under consideration, it is regarded as a single layer Markov random field (SLMRF) model, and as a double layer MRF model when the test images and classical priors are considered.

A scientific principle states that segmentation trivializes the task of bias field correction. Another principle states that the bias field distorts the intensity but not the spatial attribute of an image. This thesis exploits these two widely acknowledged scientific principles in order to propose a new model for correction of intensity inhomogeneity.

The noise estimation algorithm is invariant to the presence or absence of background features in an image and more accurate in the estimation of noise levels because it is potentially immune to the modeling errors inherent in some current state-of-the-art algorithms. The noise reduction algorithm derived from the SLMRF model does not incorporate a regularization parameter. Furthermore, it preserves edges, and its output is devoid of the blurring and ringing artifacts associated with Gaussian and wavelet based algorithms. The procedure for correction of intensity inhomogeneity does not require the computationally intensive task of estimation of the bias field map. Furthermore, there is no requirement for a digital brain atlas which will incorporate additional image processing tasks such as image registration.