Classification of disease subgroup and correlation with disease severity using magnetic resonance imaging whole-brain histograms: application to magnetization transfer ratios and multiple sclerosis

Abstract

Abstract-This paper presents a new approach to characterize subtle diffuse changes in multiple sclerosis (MS) using histograms derived from magnetization transfer ratio (MTR) images. Two major parts dominate our histogram analysis; 1) Classification of MTR histograms into control and MS subgroups; 2) Correlation with current disability, as measured by the EDSS scale (a measure of disease severity). Two data reduction schemes are used to reduce the complexity of the analysis: linear discriminant analysis (LDA) and principal component analysis (PCA). LDA is better for the classification of MTR histograms as it takes into account the between-class variation. By using LDA, the space of MTR histograms is transformed to the optimal discriminant space for a nearest mean classifier. In contrast, PCA is useful for correlation with current disability as it takes into account the variation within each subgroup in its process. A multiple regression analysis is used to evaluate the multiple correlation of those principal components with the degree of disability in MS. This is the first application of such classification and correlation techniques to magnetic resonance imaging histogram data. Our MTR histogram analysis approach give improved classification success and improved correlation compared with methods that use traditional histogram features such as peak height and peak location.Index Terms-Histogram analysis, linear discriminant analysis, multiple sclerosis, principal component analysis.

A brain magnetization transfer MRI study with a clinical follow up of about four years in patients with clinically isolated syndromes suggestive of multiple sclerosis

“…The results of the present study and the conflicting findings of previous ones may cast doubt about the actual prognostic value of MT MRI findings in CIS patients. However, the potential gain in sensitivity and predictive value, which might be achieved by using other techniques for MT MRI postprocessing and analysis, such as voxel-based morphometry [1] and principal component analysis [6], needs to be investigated by further studies of CIS. Ideally, these should be performed in large samples of patients with heterogeneous CIS presentations, include multiple quantitative MR-derived metrics and have mediumto long-term follow-up durations.…”

Sources of variation in multi-centre brain MTR histogram studies: body-coil transmission eliminates inter-centre differences

“…If the histogram shows signs of bimodality [35], fitting several peaks may be more appropriate than just characterising the largest one. Principle-component analysis and linear discriminant analysis have been shown to be effective in extracting relevant information from histograms [36], without the need for feature selection. …”

Histograms: Measuring Subtle Diffuse Disease

“…Second, correlations with disease severity (EDSS 11 in MS) have been measured, in the hope that a high correlation indicates the clinical relevance of the MR parameter. Dehmeshki has argued that a single (localized) histogram feature cannot be simultaneously optimal for these two distinct tasks, and that selecting from the range of conventional features (peak height, peak location etc., see list above) is unsatisfactory, since it leads to multiple comparisons and loss of sensitivity through the Bonferroni correction (Dehmeshki et al 2002b). Instead he proposed that LDA is optimal for maximizing the separation between groups of subjects (such as sub-groups of MS), and that it should be seen as a classification problem where the success of classifying individual subjects (rather than the separation of groups) should be the prime measure, particularly when some features can separate the groups with very low p-values (see Figure 18.9 and Tables 18.3 and 18.4).…”