The effect of acquisition resolution on orientation decoding from V1: comparison of 3T and 7T

Previously published results indicate that the accuracy of decoding visual orientation from 7 Tesla fMRI data of V1 peaks at spatial acquisition resolutions that are routinely accessible with more conventional 3 Tesla scanners. This study directly compares the decoding performance between a 3 Tesla and a 7 Tesla dataset that were acquired using the same stimulation paradigm by applying an identical analysis procedure. The results indicate that decoding models built on 3 Tesla data are comparatively impaired. Moreover, we found no evidence for a strong coupling of BOLD signal change magnitude or temporal signal to noise ratio (tSNR) with decoding performance. Direct enhancement of tSNR via multiband fMRI acquisition at the same resolution did not translate into improved decoding performance. Additional voxel selection can boost 3 Tesla decoding performance to the 7 Tesla level only at a 3 mm acquisition resolution. In both datasets the BOLD signal available for orientation decoding is spatially broadband, but, consistent with the size of the BOLD point-spread-function, decoding models at 3 Tesla utilize spatially coarser image components.

two a-priori constraints: 1) sufficient spatial coverage of the ROI, and 2) identical temporal sampling frequency (TR) across resolutions.    are presented in this study are the result of a re-analysis using the exact same procedure 178 that is described here for the 3 Tesla data. ROI sizes across field strengths (see Table 1   At 7 Tesla the average BOLD signal change in V1 increases with higher acquisition resolutions (A). However, this trend was not observed for the average signal change in the entire V1 ROI at 3 Tesla (B). Considering only "responsive" voxels, as determined by univariate feature selection (see main text for procedure), the pattern of increasing signal change observed at 7 Tesla can now be found for both field strengths, at a generally elevated level. In both cases, the signal change observed in the multiband acquisition is similar to that of the 3 mm data. Notably, the average signal change differences for the 1.4 mm vs. 2 mm acquisitions at 3 Tesla are substantially larger than those observed at 7 Tesla. The number of voxels for the comparison between panels C and D differs by a factor of 1.5-3 across field strengths (see Table 1). Panel E-F show the average BOLD signal change per orientation in the 50 respective voxels that exhibit the highest average responses to any stimulation in the whole V1 ROI, determined for each resolution and field strength separately. The observed average BOLD signal change in those voxels is similar across field in both 3T and 7T data.  indicating a higher noise level.

245
All data analyses were repeated for both field strengths using this additional feature 246 selection step. Figure 1B shows 294 We previously reported that increased tSNR does not necessarily lead to an increase The high-pass components perform better than the lowpass components beyond 5mm FWHM. The band-pass components showed peak accuracy between 5-8mm FWHM band. The peak in decoding accuracy was more pronounced after feature selection (mostly evident in 1.4mm and 2mm data). The variability in the accuracy results across subjects was higher than for the 7 Tesla data, showing the contribution of increased noise in the data. Similar patterns of decoding accuracy across different levels of smoothing were found as in the 7 Tesla data for LP, HP and BS components. The peak in 5-8mm band was not observed for band-pass components for the 1.4mm data. For the 2mm acquisition with and without parallel imaging accelaration, and the 3mm data showed a trend of increased decoding accuracy in the same band but not as pronounced as in the 7 Tesla data.
larger number of voxels in the ROI prior feature selection), and in particular in the ≈5-8 mm band ( Figure 5A-C vs. D-F).

344
The general pattern observed for decoding 3 Tesla data without feature selection 345 resembles that of the 7 Tesla data, albeit at a substantially higher noise level. LP 346 volumetric Gaussian filtering showed a general trend towards a monotonic decrease in 347 decoding accuracy with increase in kernel size, but the decoding performance remained 348 above chance-level even after spatial smoothing of 10 mm, except for the regular 2 mm 349 acquisition without feature selection that already shows a quasi-chance performance 350 without any LP filter applied ( Figure 5B). HP components showed above chance de-351 coding accuracy and performed better than the LP filtered image beyond the 9-10 mm 352 filtering kernel in all resolutions, in comparison to the ≈5 mm boundary observed for 353 7 Tesla data. Overall, peak accuracies were obtained on BP filter images.

354
In comparison to the 7 Tesla data, differences in accuracies between matching reso-355 lutions increase substantially in favor of the 7 Tesla data with higher resolutions. This