Fig. 1

Motivating LISA. The top row shows z-maps obtained by applying a voxelwise onesample t-test applied to 20 randomly generated images. The maps were computed with either no spatial filtering, or spatial filtering before versus after applying the t-test. The bottom row shows the same results thresholded so that the false discovery rate was 0.05 with the simulated signal as an underlay in light grey. Here we scaled the filtered and unfiltered z-values to the same range. Note however, that after filtering the z-values are no longer standardized so that their interpretation differs. The true positives rates were 0.035 (no filter), 0.89 (pre-test Gaussfilter), 0.97 (post-test Gaussfilter), 0.99 (post-test bilateral filter). Applying the Gaussfilter prior to the t-test diminishes spatial accuracy and yields a poor true positive rate. A better result is achieved by applying the Gaussian filter after the t-test. The post-test bilateral filter produces the best results in terms of both spatial accuracy and statistical power. Therefore, we propose this approach for our new method “LISA”