OBJECTIVES: The aims of this study were to examine whether visual MRI rating scales used in diagnostics of cognitive disorders can be estimated computationally and to compare the visual rating scales with their computed counterparts in differential diagnostics.

METHODS: A set of volumetry and voxel-based morphometry imaging biomarkers was extracted from T1-weighted and FLAIR images. A regression model was developed for estimating visual rating scale values from a combination of imaging biomarkers. We studied three visual rating scales: medial temporal lobe atrophy (MTA), global cortical atrophy (GCA), and white matter hyperintensities (WMHs) measured by the Fazekas scale. Images and visual ratings from the Amsterdam Dementia Cohort (ADC) (N = 513) were used to develop the models and cross-validate them. The PredictND (N = 672) and ADNI (N = 752) cohorts were used for independent validation to test generalizability.

RESULTS: The correlation coefficients between visual and computed rating scale values were 0.83/0.78 (MTA-left), 0.83/0.79 (MTA-right), 0.64/0.64 (GCA), and 0.76/0.75 (Fazekas) in ADC/PredictND cohorts. When performance in differential diagnostics was studied for the main types of dementia, the highest balanced accuracy, 0.75-0.86, was observed for separating different dementias from cognitively normal subjects using computed GCA. The lowest accuracy of about 0.5 for all the visual and computed scales was observed for the differentiation between Alzheimer's disease and frontotemporal lobar degeneration. Computed scales produced higher balanced accuracies than visual scales for MTA and GCA (statistically significant).

CONCLUSIONS: MTA, GCA, and WMHs can be reliably estimated automatically helping to provide consistent imaging biomarkers for diagnosing cognitive disorders, even among less experienced readers.

KEY POINTS: • Visual rating scales used in diagnostics of cognitive disorders can be estimated computationally from MRI images with intraclass correlations ranging from 0.64 (GCA) to 0.84 (MTA). • Computed scales provided high diagnostic accuracy with single-subject data (area under the receiver operating curve range, 0.84-0.94).