Volume-based bias in automated measurements of lateral ventricle and hippocampal volumes of mild traumatic brain injury patients

Lateral ventricle and hippocampal volume are clinical indicators for many disorders. Manual tracing is time-consuming; the alternative, automated tracing, may be inconsistent across software programs. The purpose was to estimate the reliability of three automated measuring programs relative to manual tracing, and to consider the use of linear inter-conversion factors. Participants with mild traumatic brain injury (mTBI) were measured for lateral ventricle volume (LVV; n = 20; 35% male; mean age = 36.5) or hippocampal volume (HV; n = 50; 42% male; mean age = 39.1). Axial 4 mm T2 images and 1.2 mm coronal T1 Images were used for manual lateral ventricle tracings and hippocampal tracings, respectively (InteleViewer®). 3T MR images 1.2 mm sagittal T1 sequences were used for automated volumetric tracings in NeuroQuant® v3.0, FreeSurfer® v6, and volBrain software. Reliability was estimated by intraclass correlation coefficients (ICCs), and relationships between software programs were described using regression equations. Human inter-rater reliability was excellent (ICC = 0.99) for LVV and moderate (0.67) for HV. For LVV, software programs had excellent consistency with manual tracing and with each other (≥0.97). There was excellent agreement between manual tracing and NeuroQuant® (0.97), but moderate-good agreement (0.54-0.87) with FreeSurfer® or VolBrain, with notably size-biased underestimation of LVV. For HV, FreeSurfer® and VolBrain had excellent consistency with manual tracing (≥0.90); NeuroQuant® had moderate-good consistency with other methods (0.57-0.79). FreeSurfer® had excellent agreement with manual tracing (0.91), but poor-good agreement with other software (0.42-0.78). Compared to manual tracing, FreeSurfer® overestimated large HV, and volBrain underestimated small HV. Conversions involving HV may be ill-advised given their low consistency. Algorithms for brain volumetrics must be evaluated for distorting effects of size-associated bias.