IC‐P‐136: Estimation of total intracranial volume; a comparison of methods

Regional cortical and subcortical grey matter and white matter volumes correlate with total brain volume, which in turn is correlated with total intracranial volume (TIV). In studies of neurodegenerative disease, total brain volume typically decreases over time, while TIV remains constant throughout adulthood. Controlling for TIV in statistical analyses can help adjust for sampling imbalances in headsize, and explain some of the variability in regional measurements without removing disease effects, thus increasing power to detect group differences or structural correlates. Time-consuming manual estimation of TIV is unsuitable for large studies such as the Alzheimer's Disease Neuroimaging Initiative (ADNI), motivating automatic methods; however, inaccurate automatic methods could exhibit disease-related biases that could confound analyses (for instance if the TIV measure was influenced by atrophy), so careful evaluation is necessary. We compare estimates of TIV using 55 subjects (37 male, 59.9 + / - 11.4 years of age), including 22 with Alzheimer's disease and 16 with frontotemporal lobar degeneration, imaged twice (479 + / - 289 days apart). Manual estimates performed with the MIDAS software are compared to eight automatic Free Surfer versions 4.5 and 5; Statistical Parametric Mapping version 5, summing either native or modulated warped segmentations; SPM version 8 equivalents; Jacobian integration using either SPM8 unified segmentation or Dartel. Considering baseline TIV, SPM5 segmentations are highly variable and upwardly biased with respect to manual measures, while those from SPM8 are dramatically improved; FreeSurfer results lie between these (figure a). The Dartel Jacobian integration method has the strongest correlation and least mean difference with respect to manual measures (see table). Consideration of longitudinal changes reveals small but significant reductions in the manual measures, and in the closely correlated Dartel results (figure b). The modulated warped SPM8 segmentations appear to exhibit the best balance of accuracy and stability over time. Accurate estimation of TIV is possible in less than 10 minutes per subject with no manual interaction, making it suitable for even the largest studies given distributed computing resources. We will use the above-selected method to estimate TIV for all subjects in ADNI, and make these publicly available. Boxplots for TIV at baseline and differences from baseline to repeat. MIDAS denotes measurements performed manually by experts using the MIDAS software. [Boxplots show the median (red line), interquartile range (box), nonparametric test for equal medians (whether angled notches overlap), and outliers (red crosses), which are further than 1.5 times the interquartile range beyond the interquartile limits (the maximum length of the boxplot whiskers).]