We propose a novel approach for evaluating the performance of activation detection in real (experimental) datasets using a new mutual information (MI)-based metric and compare its sensitivity to several existing performance metrics in both simulated and real datasets. The proposed approach is based on measuring the approximate MI between the fMRI time-series of a validation dataset and a calculated activation map (thresholded label map or continuous map) from an independent training dataset. The MI metric is used to measure the amount of information preserved during the extraction of an activation map from experimentally related fMRI time-series. The processing method that preserves maximal information between the maps and related time-series is proposed to be superior. The results on simulation datasets for multiple analysis models are consistent with the results of ROC curves, but are shown to have lower information content than for real datasets, limiting their generalizability. In real datasets for group analyses using the general linear model (GLM; FSL4 and SPM5), we show that MI values are (1) larger for groups of 15 versus 10 subjects and (2) more sensitive measures than reproducibility (for continuous maps) or Jaccard overlap metrics (for thresholded maps). We also show that (1) for an increasing fraction of nominally active voxels, both MI and false discovery rate (FDR) increase, and (2) at a fixed FDR, GLM using FSL4 tends to extract more voxels and more information than SPM5 using the default processing techniques in each package.
展开▼
