MRI visual rating scales in the diagnosis of dementia: evaluation in 184 post-mortem confirmed cases

Accurately distinguishing between different degenerative dementias during life is challenging but increasingly important with the prospect of disease-modifying therapies. Molecular biomarkers of dementia pathology are becoming available, but are not widely used in clinical practice. Conversely, structural neuroimaging is recommended in the evaluation of cognitive impairment. Visual assessment remains the primary method of scan interpretation, but in the absence of a structured approach, diagnostically relevant information may be under-utilized. This definitive, multi-centre study uses post-mortem confirmed cases as the gold standard to: (i) assess the reliability of six visual rating scales; (ii) determine their associated pattern of atrophy; (iii) compare their diagnostic value with expert scan assessment; and (iv) assess the accuracy of a machine learning approach based on multiple rating scales to predict underlying pathology. The study includes T1-weighted images acquired in three European centres from 184 individuals with histopathologically confirmed dementia (101 patients with Alzheimer's disease, 28 patients with dementia with Lewy bodies, 55 patients with frontotemporal lobar degeneration), and scans from 73 healthy controls. Six visual rating scales (medial temporal, posterior, anterior temporal, orbito-frontal, anterior cingulate and fronto-insula) were applied to 257 scans (two raters), and to a subset of 80 scans (three raters). Six experts also provided a diagnosis based on unstructured assessment of the 80-scan subset. The reliability and time taken to apply each scale was evaluated. Voxel-based morphometry was used to explore the relationship between each rating scale and the pattern of grey matter volume loss. Additionally, the performance of each scale to predict dementia pathology both individually and in combination was evaluated using a support vector classifier, which was compared with expert scan assessment to estimate clinical value. Reliability of scan assessment was generally good (intraclass correlation coefficient > 0.7), and average time to apply all six scales was <3 min. There was a very close association between the pattern of grey matter loss and the regions of interest each scale was designed to assess. Using automated classification based on all six rating scales, the accuracy (estimated using the area under the receiver-operator curves) for distinguishing each pathological group from controls ranged from 0.86-0.97; and from one another, 0.75-0.92. These results were substantially better than the accuracy of any single scale, at least as good as expert reads, and comparable to previous studies using molecular biomarkers. Visual rating scores from magnetic resonance images routinely acquired as part of the investigation of dementias, offer a practical, inexpensive means of improving diagnostic accuracy.