Minimizing biases in estimating the reorganization of human visual areas with BOLD retinotopic mapping

Abstract

There is substantial interest in using functional magnetic resonance imaging (fMRI) retinotopic mapping techniques to examine reorganization of the occipital cortex after vision loss in humans and nonhuman primates. However, previous reports suggest that standard phase encoding and the more recent population Receptive Field (pRF) techniques give biased estimates of retinotopic maps near the boundaries of retinal or cortical scotomas. Here we examine the sources of this bias and show how it can be minimized with a simple modification of the pRF method. In normally sighted subjects, we measured fMRI responses to a stimulus simulating a foveal scotoma; we found that unbiased retinotopic map estimates can be obtained in early visual areas, as long as the pRF fitting algorithm takes the scotoma into account and a randomized "multifocal" stimulus sequence is used.

Keywords: cortical reorganization; fMRI; partial visual loss; retinotopic mapping; scotoma.

Figure 1

Stimuli. Panels A and B show an individual frame of the multifocal and the drifting bars stimulus sequences in the full-field condition (arrows in panel B and D indicate the direction of bar motion). Panels C and D show the same frame for the scotoma condition, with the yellow dashed circle (not part of the display) marking the extent of the mean luminance mask simulating a foveal scotoma.

Figure 2

The components of the pRF linear model. The input stimulus and the pRF are represented as matrices, where vertical and horizontal space are collapsed into a single dimension (px = 1…n) and time is the other dimension (t = 1…m).

Figure 3

Goodness of fit of pRFs in the full-field condition: distributions (data pooled across subjects) and brain maps for one exemplificative subject. The blue line in panels A, B, and E outlines the surface projection of the ROI selected for pRF fitting; color shading represents the goodness of fit of pRF estimates for the multifocal (A) and the drifting bars stimulation sequence (B), mapping only voxels that were successfully fit (goodness of fit >0.4 and other criteria, see Methods). Panel C and D plot the proportion of successfully fit voxels in each of the V1–V3 ROIs (stacked bars for different goodness-of-fit indices, same color coding as in panels A and B). The yellow area in panel E marks voxels for which pRFs were successfully fit using both stimulation sequences and were considered for the analysis in Panel F of this figure (showing the normalized histograms of goodness-of-fit values) and in Figures 4 and 6.

Figure 4

pRFs in the full-field condition for the multifocal and the drifting bars stimulation sequences. pRF size is plotted against pRF eccentricity; voxels are binned in contiguous 0.5° steps. Symbols show medians in each bin for the two stimulus sequences. Continuous lines are linear fits through the data points in each series, weighted by the inverse of their SEM (error bars, smaller than symbol size). This analysis included all voxels successfully fit using both the multifocal and the drifting bars sequence.

Figure 5

Schematic illustration of our simulations. pRFs estimated from the full-field condition (three examples in panels A–C, top) and the stimulus sequence from the scotoma condition were entered in the “forward solution” to obtain simulated fMRI responses (black lines and circles). For the purpose of this illustration, the time dimension is ignored; stimulus and response are plotted as a function of eccentricity. Two methods were then used to recover the pRF parameters from the simulated fMRI responses: the full-stimulus pRF method, where the full-field stimulus was used as input to the pRF fitting procedure, and the effective-stimulus pRF method, where the scotoma is modeled in the stimulus representation. The bottom rows show the recovered pRFs with the two methods (red and green, respectively) together with the original pRFs (blue).

Figure 6

Comparison of pRFs from the full-field and the scotoma condition. Separate panels show results for the multifocal (A and C) and drifting bars stimulus sequence (B and D). The gray shadow represents the scotoma region. Panels A and B plot the eccentricity of pRFs from the scotoma condition against the eccentricity of pRFs from the full-field condition. Inset bar graphs show the median eccentricity difference between pRF eccentricity in the scotoma and full-field conditions (voxels with full-field pRFs between 1.5° and 2.5°; ***p < 0.001). Panels C and D plot the size of pRFs obtained in each condition and with each analysis method against the respective eccentricity. Voxels were pooled across subjects and binned into contiguous 0.5° bins; symbols report medians in each bin and error bars show SEM.

Figure 7

Comparison of predicted and observed pRFs shifts across ROIs. Bar graphs adopt the same format as the inset in Figure 6A and B, showing the median eccentricity difference between pRF eccentricity in the scotoma and full-field conditions (voxels with full-field pRFs between 1.5° and 2.5°), for the simulations and the pRF estimates obtained with the full-stimulus pRF method. For each ROI, parentheses give the number of voxels included in the analysis in each ROI and for each stimulus sequence (this is different than in Figures 4 and 6, where the sample of analyzed voxels was the same for the two stimulus sequences).