Fig 1.
a–tSNR map from the same volunteer, scan and sagittal slice after covSoS, STARC, STARCps and STARCtsm coil combinations. STARC yielded the highest tSNR map. b–Ratio of STARC and covSoS tSNR maps after median image filtering. STARC always improves tSNR, up to a factor 2. c–tSNR distributions pooled across all volunteers. Overall, STARC outperforms the other coil combinations in terms of tSNR. A brain mask was used to ignore non brain voxels.
Fig 2.
a–Activation maps for the motor contrast at p<0.001 (no correction) for covSoS, STARC, STARCps and STARCtsm. No double-dipping correction was applied to STARCtsm. The shown maps are from the same volunteer as in Fig 1. b–Total number of activated voxels relative to the total number of voxels at different p-values and for different coil combinations, pooled over the volunteers. STARCps and STARC have the poorest performance in terms of BOLD detection. CovSoS has the highest number of activations for the lowest p-values but STARCtsm (without double-dipping correction) reports more activations for the highest p-values.
Fig 3.
t-score gains versus tSNR or t-score for the mean gains.
a—Scatter plot linking the activated voxels, the tSNR gain with STARC to its t-score gain for motor contrast. b–Scatter plot linking the t-score for the mean gain with STARC to its t-score gain for motor contrast. c–Same plot than a but with auditory contrast. d–Same plot than b but with auditory contrast. Each point corresponds to an activated voxel (p<0.001) according to the covSoS combination maps. In general, an increase of tSNR did not yield a better t-score.
Fig 4.
a–T-score distributions of covSoS and STARCtsm from noise signals and Monte Carlo simulations. The distribution of STARCtsm under the null hypothesis is distorted compared to covSoS because of double-dipping. b–Total number of activations relative to the total number of voxels at different p-values and for covSoS, STARCtsm and its double dipping corrected version for the in vivo scan. After correction, the number of activations from STARCtsm markedly drops. c–Activation maps on one volunteer for visual vs auditory contrast at p<0.001 (no correction). After double-dipping correction, activations are weaker and clusters smaller.
Fig 5.
Weights distributions from STARC optimization for two voxels.
a–Brain image from volunteer #2 axial view, the red points labelled A and B are the voxels whose weights are analysed. b–For both voxels A and B, the bar plots display the temporal mean, the temporal standard deviation and the optimized STARC weights. The channels receiving the strongest signal have the highest variability but will be given the lowest weight. c–Scatter plot of the mean versus the standard deviation for each channel, the slope of the linear fit of the ten strongest channels is also plotted. The signal variability is highly proportional to the signal strength.
Fig 6.
Signal time series for two activated voxels for covSoS and STARC.
A time series of the stimuli onset convolved with the canonical hrf is also displayed. Each graph corresponds to a voxel. CovSoS has higher activation peaks than STARC but the latter has the highest tSNR.