Evaluation of outliers in acquired brain MR images

Abstract

Pre-processing is an important stage in the analysis of magnetic resonance images (MRI), because the effect of specific image artefacts, such as intensity inhomogeneity, noise and low contrast can adversely affect the quantitative image analysis. The image histogram is a useful tool in the analysis of MR images given that it allows a close relationship with important image features such as contrast and noise. The noise and variable contrast are elements that locally modify the quality of images. The key issue of this study derives from the fact that the spatial histogram can contain outliers indicating corrupted image information through the disorder of the bins. These aberrant errors should be excluded from the studied data sets. Here, the outliers are evaluated by using rigorous methods based on the probability theory and Chauvenet (CC), Grubbs (GC) and Peirce's (PC) criteria. In order to check the quality of the MR images, the Minkowsky (MD), Euclidean (ED) and cosine (CD) distance functions were used. They act as similarity scores between the histogram of the acquired MRI and the processed image. This analysis is necessary because, sometimes, the distance function exceeds the co-domain because of the outliers. In this paper, 32 MRIs are tested and the outliers are removed so that the distance functions generate uncorrupted and real values.