COMPUTED RATING SCALES FOR COGNITIVE DISORDERS FROM MRI

Abstract

Magnetic resonance imaging (MRI) is an important tool in the clinical diagnosis of cognitive disorders. In addition to visual reading of images, visual rating scales are used to evaluate images in clinical practice. The objective of this study is to assess whether widely used visual rating scales, medial temporal lobe atrophy (MTA)1, global cortical atrophy (GCA)2 and white matter hyperintensities (WMHs) measured by the Fazekas scale3, could be estimated automatically. Images and visual ratings from the Amsterdam Dementia Cohort4 (Table 1) were used to develop the models. A set of imaging biomarkers were extracted from T1-weighted and FLAIR images using methods similar to 5, and corrected for age, gender and head size. The biomarkers tested included: volumes for hippocampus (MTA), inferior lateral ventricle (MTA), cortical gray matter (GCA), global WMHs (Fazekas) and deep WMHs (Fazekas), and gray matter concentrations from voxel-based morphometry6 for hippocampus (MTA) and cortex (GCA). A regression model was developed for estimating the visual rating scale values from the imaging biomarkers. Cross-validation was used. The Mann-Whitney U-test was used to test statistical differences. Correlation coefficient between the visual and computed rating scale values was 0.86 (MTA-Left), 0.85 (MTA-Right), 0.67 (GCA) and 0.80 (WMHs). In classifying cognitively normal subjects and Alzheimer's disease patients, the area under the curve (AUC) for the visual scales was 0.82 (MTA-left), 0.79 (MTA-right) and 0.76 (GCA), and for the computed scales 0.88 (MTA-Left), 0.89 (MTA-Right) and 0.90 (GCA). The corresponding AUC values for classifying cases with vascular cognitive impairment and other etiologies were 0.88 (visual) and 0.96 (computed). The differences between the visual and computed scales were statistically significant. Figure 1 shows ROC curves for the left MTA, GCA and Fazekas. MTA, GCA and WMHs can be reliably measured automatically. The computed ratings performed better than current visual ratings in classification. Age correction was not used for visual ratings which may partly explain the result. 1 Scheltens J Neurol 1995; 2 Pasquier Eur Neurol 1996; 3 Fazekas Am J Roentgenol 1987; 6 van der Flier J Alzheimers Dis 2014; 5 Koikkalainen NeuroImage Clinical 2016; 6 Ashburner NeuroImage 2000. ROC curves for the visual and computed MTA-L, GCA and Fazekas rating scales. Two curves are shown for the computed scales: the original scalar scale and the scale rounded to an integer value corresponding to the range of values in the visual scales.