Quantification of retinotopic maps with a Gaussian process modeling

Visual neuroscientists use blood oxygenation level-dependent fMRI data to delineate the structure and boundaries of visual-field representations on the human cortex, a process called retinotopic mapping. Although the population receptive field (PRF) model facilitates receptive-field estimation for cortical voxels, retinotopic map quantification still faces challenges. In vivo, retinotopic areas exhibit consistent topology, but modeling this topology is often disrupted by limited resolution and low signal-to-noise ratio. Additionally, automated segregation of visual areas, including the fovea, is lacking. We address these challenges with three innovations: (1) extended polar angle parametrization, (2) cortical anchor point identification, and (3) map estimation via a Gaussian process model. The Gaussian process provided significantly improved estimated generalization error than linear regression and reduced topological violations in estimated maps from 49.2% to 31.5% along the total analyzed cortical mesh from the left hemispheres of 181 subjects from the Human Connectome Project. It automatically defined precise boundaries between the six discrete visual areas, along which the mean 95% credible interval width was 0.104 π rad. Along the estimated eccentricity contour of 3°, the mean 95% credible interval width was 0.586°. We estimated the foveal confluence location from fMRI to be systematically more dorsal and medial than the occipital pole across all subjects. This Gaussian process modeling approach offers a more accurate and reliable method for quantifying retinotopic maps.