Concordance between computer-based neuroimaging findings and expert assessments in dementia grading

Abstract

We present computer-based techniques for the measurement, detection and quantification of dementia related neuroimaging findings. The three focal points of these are: brain atrophy, lateral ventricle enlargement, and white matter hyperintensity (WMH) lesion detection and quantification. Research has shown that pathological changes in these three areas have strong correlations with dementia. Our proposed methods seek to provide consistent and accurate measures of these pathologies by applying advanced image processing techniques to brain MRIs. The results are evaluated by investigating correlations with experts' grades. We report correlations of up to 0.79, 0.70 and -0.24 for hemispheric cortical atrophy, central sulcus width, and central sulcus depth, respectively. For lateral ventricle (LV) enlargement, we obtain an LV-to-brain ratio correlation of up to 0.81 among other measurements. We automatically segment and quantify periventricular WMH lesions by adaptive thresholding and 3D connectivity analysis on fluid attenuation inversion recovery MRIs. A mean Dice score of 0.83 (standard deviation is 0.06), and a correlation score of up to 0.84 is obtained.