An evaluation of Z-transform algorithms for identifying subject-specific abnormalities in neuroimaging data

The need for algorithms that capture subject-specific abnormalities (SSA) in neuroimaging data is increasingly recognized across many neuropsychiatric disorders. However, the effects of initial distributional properties (e.g., normal versus non-normally distributed data), sample size, and typical preprocessing steps (spatial normalization, blurring kernel and minimal cluster requirements) on SSA remain poorly understood. The current study evaluated the performance of several commonly used z-transform algorithms [leave-one-out (LOO); independent sample (IDS); Enhanced Z-score Microstructural Assessment of Pathology (EZ-MAP); distribution-corrected z-scores (DisCo-Z); and robust z-scores (ROB-Z)] for identifying SSA using simulated and diffusion tensor imaging data from healthy controls (N = 50). Results indicated that all methods (LOO, IDS, EZ-MAP and DisCo-Z) with the exception of the ROB-Z eliminated spurious differences that are present across artificially created groups following a standard z-transform. However, LOO and IDS consistently overestimated the true number of extrema (i.e., SSA) across all sample sizes and distributions. The EZ-MAP and DisCo-Z algorithms more accurately estimated extrema across most distributions and sample sizes, with the exception of skewed distributions. DTI results indicated that registration algorithm (linear versus non-linear) and blurring kernel size differentially affected the number of extrema in positive versus negative tails. Increasing the blurring kernel size increased the number of extrema, although this effect was much more prominent when a minimum cluster volume was applied to the data. In summary, current results highlight the need to statistically compare the frequency of SSA in control samples or to develop appropriate confidence intervals for patient data.