Quantile estimation to derive optimized test thresholds for random field statistics

Statistical framework and noise sensitivity of the amplitude radial correlation contrast method

A statistical framework for the amplitude radial correlation contrast (RCC) method, which integrates a conventional pixel threshold approach with cluster-size statistics, is presented. The RCC method uses functional MRI (fMRI) data to group neighboring voxels in terms of their degree of temporal cross correlation and compares coherences in different brain states (e.g., stimulation OFF vs. ON). By defining the RCC correlation map as the difference between two RCC images, the map distribution of two OFF states is shown to be normal, enabling the definition of the pixel cutoff. The empirical cluster-size null distribution obtained after the application of the pixel cutoff is used to define a cluster-size cutoff that allows 5% false positives. Assuming that the fMRI signal equals the task-induced response plus noise, an analytical expression of amplitude-RCC dependency on noise is obtained and used to define the pixel threshold. In vivo and ex vivo data obtained during rat forepaw electric stimulation are used to fine-tune this threshold. Calculating the spatial coherences within in vivo and ex vivo images shows enhanced coherence in the in vivo data, but no dependency on the anesthesia method, magnetic field strength, or depth of anesthesia, strengthening the generality of the proposed cutoffs.

An improved theoretical P value for SPMs based on discrete local maxima

We present a new continuity correction to the P value for local maxima of a statistical parametric map that bridges the gap between small FWHM, when the Bonferroni correction is accurate, and large FWHM, when random field theory is accurate. The new method, based on discrete local maxima, is always an upper bound (like the Bonferroni), but lower and hence more accurate for large FWHM, without increasing false positives. It resulted in P values that were approximately 43% lower than the best of Bonferroni or random field theory methods when applied to a typical fMRI data set.