Reliability of PET activation across statistical methods, subject groups, and sample sizes.

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Medical Imaging Analysis

Background:

• Positron emission tomography (PET) is crucial for visualizing brain activity.
• Accurate estimation of regional activation is vital for interpreting PET studies.
• Various pixel-based statistical methods exist for PET activation analysis, each with potential strengths and weaknesses.

Purpose of the Study:

• To compare the performance of four distinct pixel-based methods for estimating regional activation in PET images.
• To evaluate the impact of technical factors on method performance under controlled conditions.
• To assess the replicability and reliability of activation detection across different methods.

Main Methods:

• Implementation and comparison of four pixel-based methods: change distribution analysis, Worsley's method, a pixelwise general linear model, and a nonparametric method.
• Utilized a large dataset from a verb generation paradigm with 18 normal subjects.
• Held key technical factors constant, including smoothing, stereotactic transform, coregistration, search volume, and alpha level.

Main Results:

• All methods controlled Type I errors (false positives) at the nominal rate in noise datasets.
• Detected regions of activation demonstrated high internal replication (93%).
• Methods differed primarily in Type II error rates and the stability of activation cluster locations; local variance-based methods were less powerful and stable with small sample sizes.

Conclusions:

• Worsley's method using pooled variance improved stability but faced challenges with non-stationary variance.
• The power of current pixel-based analyses is modest with conventional sample sizes.
• Modeling variance sources and improving anatomical standardization are key to enhancing the power of PET activation analyses.