Ventricle shape features as a reliable differentiator between the behavioral variant frontotemporal dementia and other dementias

Abstract

AbstractBackgroundIn the absence of a pathologic genetic mutation, diagnostic certainty of the behavioral variant frontotemporal dementia (bvFTD) still relies on convergence of clinical criteria and imaging findings. Using deformation‐based morphometry (DBM), we recently showed that ventricular volume can discriminate bvFTD from cognitively normal controls (CN)1. Here, we investigate the performance of shape‐based ventricular features to differentiate bvFTD from Alzheimer’s Dementia (AD), mild cognitive impairment (MCI), semantic and progressive non‐fluent aphasia variants of FTD (SV and PNFA), and CN.MethodData included 825 participants: 59 bvFTD, 28 SV, 30 PNFA, and 105 CN from the Frontotemporal Lobar Degeneration Neuroimaging Initiative (FTLDNI) and 322 age‐matched amyloid β+ MCI, 127 age‐matched amyloid β+ AD and 164 age‐matched CN from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). A previously validated patch‐based label fusion technique was employed to segment the lateral ventricles 2. Linearly registering the images to ICBM2009c template, the ventricles were divided into anterior‐posterior (using coronal slice y=120) and left‐right halves (separately for frontal and temporal lobes). Total ventricle volume (TVV), anterior‐to‐posterior ratio (APR) and frontal and temporal parenchymal left‐right ratios (LRFR and LRTR) were used alone and in combination with each other, age, and sex, to differentiate bvFTD from all other cohorts. A support vector machine classifier (fitcsvm from MATLAB with a linear kernel) was trained on each feature set to perform the classification task using 10‐fold cross validation, repeated 100 times.ResultUsing APR to identify bvFTD from a mixed age‐matched cohort (Control, MCI, AD, SV and PNFA) yielded an accuracy of 92%. Adding additional features did not improve global classification performances. The top accuracies against each individual cohort were 89% for bvFTD vs Control, 91% for bvFTD vs MCI using APR+TVV, and 83% for bvFTD vs AD using APR+LRTR. The best accuracies discriminating bvFTD from SV and PNFA were 71% and 76% respectively, using APR+TVV+LRTR+LRFR.ConclusionThe APR is an easy to obtain and generalizable ventricle‐based feature from T1‐weighted MRIs (routinely acquired and available in the clinic) that can be used to differentiate bvFTD from normal subjects, other FTD variants, MCI, and AD patients.