Robust MRI brain tissue parameter estimation by multistage outlier rejection

Abstract

This article addresses the problem of the tissue type parameter estimation in brain MRI in the presence of partial volume effects. Automatic MRI brain tissue classification is hampered by partial volume effects that are caused by the finite resolution of the acquisition process. Due to this effect intensity distributions in brain MRI cannot be well modeled by a simple mixture of Gaussians and therefore more complex models have been developed. Unfortunately, these models do not seem to be robust enough for clinical conditions, as the quality of the tissue classification decreases rapidly with the image quality. Also, the application of these methods for pathological images with unmodeled intensities (e.g. MS plaques, tumors, etc.) remains uncertain. In the present work a new robust method for brain tissue characterization is presented, treating the partial volume affected voxels as outliers of the pure tissue distributions. The proposed method estimates the tissue characteristics from a reduced set of intensities belonging to a particular pure tissue class. This reduced set is selected by using a trimming procedure based on local gradient information and distributional data. This feature makes the method highly tolerant of a large amount of unexpected intensities without degrading its performance. The proposed method has been evaluated using both In medical imaging a single voxel may be composed of a mixture of tissue types due to the finite spatial resolution of imaging devices. This phenomenon is known as the partial volume effect (PVE) and it is mainly observable at tissue type boundaries. PVE is an important factor in the study of small brain structures or highly convoluted brain regions such as those within the cerebral cortex. In fact, ignoring this effect can lead to volume measurement errors in the range of 20 -60% (1). In addition to complicating MRI-based morphometry, PVE causes severe difficulties in modeling intensities of different tissue types. For example, modeling the histogram by a simple mixture of Gaussians (MoG) does not work properly because a large part of the voxels is affected by PVE. These complications have motivated the development of more realistic intensity models and parameter estimation schemes taking the PVE into account.Most of the PVE modeling methods model both pure tissue and mixed voxels using different intensity distributions. One of the first attempts to statistical partial volume (PV) modeling was proposed by Santago and Gage (2). They modeled pure tissues assuming equal variance Gaussian distributions (motivated by the MR physics) while PV voxel intensities followed a different distribution derived based on the parameters of the pure distributions. Laidlaw et al. (3) used a similar model but incorporated a more complex method that took into account neighborhood voxels. Ruan et al. (4) concluded that PV voxels can be generally modeled by using Gaussian distributions. However, this approach has the drawback of not being able to compute PV fractions, since there is no explici...

Improved estimates of partial volume coefficients from noisy brain MRI using spatial context

“…As reference, we used the Trimmed Minimum Covariance Determinant method (TMCD) (Tohka et al, 2004) and the Trimmed Mean Segmentation (TMS) method (Manjón et al, 2008). The TMS method does not supply PVCs but the mean intensities of different brain tissue types.…”

“…In the latter work, two methods known as Minimum Covariance Determinant (MCD) and Minimum Volume Ellipsoid (MVE) were used over a set of pre-classified and trimmed data. Recently, an extension of this method has been developed (Manjón et al, 2008) that improves the tissue parameter estimation step by applying a new trimming procedure based on local gradient information to select pure tissue samples from the data.…”

“…The main aim of the present work is to combine robust parameter estimation methods (Tohka et al, 2004;Manjón et al, 2008) to strategies to reduce random noise present in MR images to formulate an accurate method to estimate the PVCs. Concretely, to reduce the contribution of image noise, the applications of a novel MRF type strategy capable to overcome the limitations of the MRF-type methods described above and a state-of-art adaptive filtering method based on non-local means modelling are studied and compared.…”

FAST-PVE: Extremely Fast Markov Random Field Based Brain MRI Tissue Classification

“…In particular, we use the maximum covariance determinant (MCD) method (also termed least trimmed squares (LTS) in one dimension) for the task [15]. The accuracy of this kind of parameter estimation has been confirmed in our earlier studies [7,16] and the reasons for selecting MCD as the parameter estimation method (instead of, e.g., more conventional M-estimators) has been explained in detail in [7], where the method just described is abbreviated as trimmed MCD (TMCD).…”

“…The initial labeling is generated by the incremental k-means technique proposed in [16] which is very fast and accurate enough. The algorithm itself is a specific multistart adaptation of the standard k-means algorithm.…”