Single-shot multi-echo EPI acquisition methods have received much attention in fMRI and been proposed to provide advantages over the single-echo variant¹. Recently, Poser and colleagues showed several approaches for combining multi-echo datasets and found that contrast-to-noise-ratio (CNR) weighting provided the highest BOLD sensitivity². Here we further investigated the behaviour of different echo weighting approaches based on fMRI time series data with two echo times both in resting-state and task-based experiments. Our first aim was to investigate whether the tendency of sensitivity increase in resting-state found by Poser et al. in multi-echo EPI holds for double-echo acquisition. In addition, a recent study by Kirilina and colleagues compared the CNR weighting for multi-echo combination with other advanced acquisition methods and reported that at the group level neither method offered major benefits compared to the simple single echo acquisiton³. However, they did not compare their results to other possible echo combination methods. Therefore here we performed group-level random-effects analyses on double-echo data using a variety of echo weightings and present a quantitative analysis of the results.All experiments were performed on a 3T Philips scanner involving 27 healthy subjects. Each participant gave written informed consent and the study was approved by local ethical comittee. 36 axial 2.6 mm slices of single-shot double-echo GE-EPI data were collected with 2-fold SENSE acceleration, TE₁/TE₂=17/44 ms, TR=2.6 s, FOV 200x200 mm², inplane resolution 2.5x2.5 mm², slice gap 0.6 mm. Resting-state data contained 116 TRs, task data contained 282 TRs. During the task experiment subjects played a competitive game against a human opponent in which the parametric modulators represented reward. Preprocessing and statistical analysis were done using SPM12⁴ and Matlab (The MathWorks, MA, USA). All volumes were realigned and detrended with voxelvise second-order polynomial fit, then linearly combined in four ways using voxelwise weights w₁ and w₂ for the data with TE₁ and TE₂ as shown in Figure 1a. Both the combined resting-state and task-driven datasets were coregistered to the corresponding anatomical images and normalized to MNI space. The resting-state data were kept unsmoothed while the task datasets were smoothed with a 6 mm isotropic Gaussian kernel. CNR calculation was performed on normalized resting-state data as described in Figure 1b. For the resting-state data average and standard deviation of CNR were calculated across subjects. For task data standard SPM statistical analysis was performed on the normalized images of each of the four echo combinations independently. For quantitative comparison of different echo combinations a map was calculated from the four t-value maps that voxelwise shows which echo combination resulted in the highest t-value in a given voxel. In addition, scatter plots were produced indicating the relationship between the group-level t-values of the simple echo averaging vs. the other three weighting approaches on a voxel-by-voxel level.In resting-state data the usage of advanced echo weighting methods clearly increases the mean CNR as shown in Figure 2a. In particular we confirm the finding of Poser et al. that the CNR weighting is the optimal echo combination. However, Figure 2b indicates a similar tendency in the standard deviation of CNR across subjects, forecasting that a group-level analysis may not benefit from the increased bold sensitivity of the CNR-weighted echo combination. Figure 3a shows the results of the group-level random effects analysis in which the echoes were simply averaged. The group-level results were highly similar for the other three echo weighting schemes (data not shown). Correspondingly, neither of the four combination methods seems to indicate a clear advantage, not even in the orbitofrontal cortex, as can be seen from Figure 3b where the color code in each voxel indicates the echo combination with the largest t-value.This result is confirmed by the results plotted in Figure 4. The t-value points are generally close to the identity line which again implies no substantial differences between methods.We quantitatively compared the performance of different echo combination methods in dual-echo fMRI analysis and showed that benefits of advanced weightings existing in single-subject level disappear in group analyses, as it was suspected from the increase of inter-subject variance of BOLD contrast with these combinations. From our results it may be concluded that the traditional averaging approaches for a multi-echo group experiment suffice, while saving significant scanner time by omitting the resting-state measurements which would otherwise be necessary for weight calculation. Future work will investigate the relation of inter- and intra-subject variance components across weighting approaches.