Computational protocol: Visual Population Receptive Fields in People with Schizophrenia Have Reduced Inhibitory Surrounds

Similar protocols

Protocol publication

[…] All functional data were preprocessed using SPM8 (http://www.fil.ion.ucl.ac.uk). All functional images were intensity bias corrected using in-house software to aid automated preprocessing of the images. The dummy volumes were then discarded and the remaining images from the mapping and HRF scans were realigned and unwarped (using the B0 field maps to correct any image distortion) and coregistered to the individual's high-resolution T1 structural image acquired with the coil on, using the additional MPRAGE structural image acquired with the front of the head coil off as an intermediate step.Freesurfer software (http://surfer.nmr.mgh.harvard.edu, version 5.0.0) was used to create 3D surface meshes of both cortical hemisphere for each individual, one for the boundary between gray and white matter and one for the outer pial boundary of the white matter. The cortical surfaces were then inflated.All further analyses were performed using a custom MATLAB toolbox developed in-house (http://dx.doi.org/10.6084/m9.figshare.1344765) for pRF analysis and for projecting data onto the cortical surface. Data analysis was restricted to a region including the occipital, posterior temporal, and posterior parietal areas defined manually for each individual. To project functional data onto the smoothed gray/white matter surface, we determined the point midway between the gray/white and pial surfaces for each vertex on the gray/white matter boundary and used this gray matter voxel to create a functional time series for each vertex for all mapping and HRF scan runs. Linear detrending and z-score normalization were applied to these time series.To estimate each individual's HRF, we averaged the signal evoked by the 10 photic bursts of the HRF scan. Outliers greater than ±1.5 SDs from the mean were excluded from the time series of each vertex. Analysis was restricted to only visually active vertices, defined by a response >1 SEM averaged over the first 5 scans after each burst. A double-gamma function was then fitted to the averaged stimulus evoked response to estimate the HRF for each hemisphere independently. There were four free parameters: the latency of the peak response and the undershoot, the peak amplitude, and the ratio of the peak and undershoot amplitudes.For the pRF analysis, we used a forward modeling approach similar to that described by to estimate the pRF parameters for each vertex independently. The pRF was initially modeled as a 2D Gaussian in visual space with four free parameters: x and y describe the pRF center position relative to the fixation point; σ (σ1) denotes the SD of the Gaussian, reflecting the spatial spread of the pRF (i.e., pRF size); and β (β1) is the response amplitude at x, y. In a subsequent analysis, we used a DoG model (based on ) that incorporated an inhibitory surround in addition to the excitatory center. Because the DoG model is described by a combination of two Gaussians (a central positive isotropic Gaussian and a second larger negative isotropic Gaussian), there are two additional parameters in the model fit: the SD of the larger negative surround (σ2) (i.e., pRF surround size) and the amplitude ratio of the two Gaussians (β2/β1).A linear overlap between the pRF model and a binary mask of the stimulus across time was used to predict the response of the neuronal population at each vertex. This predicted neuronal response was then convolved with each individual's specific HRF before optimization of the fit between this predicted neuronal response and the measured BOLD responses.We ran a first pass coarse fit on heavily smoothed functional data [Gaussian kernel with full-width at half-maximum (FWHM) = 8.3 mm]. Using a 3D search space comprising 15 × 15 × 34 combinations of location (x, y) and pRF size (σ), we calculated the Pearson correlation between the time series at each vertex and the search grid to find the parameters with the highest correlation between observed and predicted time series (because it is based on correlation the coarse fit did not include the β parameter). All vertices in the defined occipital area were included in this initial model fit. However, vertices for which the goodness-of-fit (R2) failed to reach 0.05 in the initial coarse fit were not analyzed further. The coarse-fit parameters were then used to seed a subsequent optimization process to fit the pRF parameters to unsmoothed data at each vertex by minimizing the residual sum of squares between the predicted and observed time series. This model-fitting stage included the β parameter. We also used the coarse-fitting parameters from the standard Gaussian model to seed the optimization procedure for the DoG model. Finally, we applied a surface based smoothing kernel of 5 mm FWHM to deal with any gaps in the maps arising at vertices with poor model fits. This is particularly important for the calculation of cortical surface area and the area subtended by each face in the surface mesh in visual space. See for additional details of the model-fitting procedure.Visual regions were delineated manually in Freesurfer by displaying pseudocolor-coded maps of polar angle and eccentricity calculated from the pRF analysis (). Visual areas V1–V3 were defined using standard criteria (; ; ) and V4 was defined as a full hemifield representation adjacent to the ventral portion of V3 ().To calculate the cortical magnification factor (CMF) (), we divided the square root of visual area (as determined by the distances of each pRF to the pRF positions of its cortical neighbors) by the corresponding square root of the cortical surface area calculated in the same way. To measure the macroscopic surface area of these regions, we summed the area estimates of all vertices with pRF locations that fell between 2° and 7° to avoid edge artifacts.For the pRF size data (σ), we subdivided the vertices into 6 1° wide eccentricity bins between 1.5° and 7.5°, thus avoiding the innermost and outermost pRFs, which were only partially mapped. For each individual, we calculated the mean pRF size (σ) for each eccentricity bin within each visual area V1–V4. We then calculated the group mean (σ) for each bin in all visual areas and fitted a linear regression to the data (). Any data points >2 SDs from the mean were considered outliers and were removed from the group analysis. There was no significant effect of eccentricity, visual area, pRF model, or group on the number of outliers.For the DoG model, the Gaussian with the larger SD (the negative surround) was subtracted from the smaller Gaussian (the positive center). This results in a change to the effective positive pRF size because the width (size) of the DoG excitatory component results from a combination of the center and surround parameters and their amplitude ratio. Therefore, to allow direct comparison of the excitatory components of the two models, we followed the method of and calculated the FWHM, which measures the width of the positive Gaussian at half the maximum response level. For the size of the inhibitory component of the pRF, we used the SD (σ2) of the negative-surround Gaussian from the DoG model.To compare between groups, we calculated the difference in (squared) area under the curve (linear regression) fitted to the pRF size data (FWHM or σ) plotted against eccentricity (). To confirm that observed differences were robust, we bootstrapped the fit by resampling each group 1000 times (with replacement), refitted the curves, and recalculated the difference in squared area under the curve for each iteration. The proportion of bootstrapped differences that were opposite to the observed difference was calculated. All probability values were then corrected for multiple comparisons using the false discovery rate (FDR) with a threshold of q = 0.05. […]

Pipeline specifications

Software tools SPM, FreeSurfer
Application Magnetic resonance imaging
Organisms Homo sapiens