Siddharth Mittal¹ (), Michael Woletz¹, David Linhardt¹, Christian Windischberger¹; ¹Center for Medical Physics and Biomedical Engineering, Medical University of Vienna

Population receptive field (pRF) mapping quantifies retinotopic organisation by estimating receptive-field location and size across the visual cortex. Most approaches use a two-stage fitting process: gridfit to obtain an initial solution, followed by refinefit to optimise pRF parameters. Despite its widespread use, the actual impact of refinefit and whether the improvement justifies the computation time has not been quantified systematically. We evaluated refinefit adjustments using two software implementations: mrVista (Dumoulin & Wandell, 2008) and GEM-pRF (Mittal et al., 2025). Both use similar modelling, but GEM-pRF applies a reformulated mathematics tailored for GPU computation and improved computational performance. Analyses were performed on the NYU retinotopy dataset (44 subjects; Himmelberg et al., 2021), focusing on V1. The stimulus radius was 12.4°, and both analyses used a 24.8° search space. mrVista used a triangular grid with 0.25° spatial spacing and hybrid pRF-size spacing, whereas GEM-pRF used a Cartesian grid with 0.33° spatial and 0.3° size spacing. For each voxel, gridfit and refinefit estimates of pRF centre (x, y) and size (σ) were extracted, including only voxels with >10% variance explained. In mrVista, 46.18% of voxels and in GEM-pRF, 11.97% voxels, showed no change between gridfit and refinefit. Among voxels that did change, 75.53% in GEM-pRF and 72.09% in mrVista stayed closer to their original gridfit position than to any neighbouring grid point. Overall, 78.46% (GEM-pRF) and 84.98% (mrVista) either remained on a grid point or shifted by less than the grid spacing. These adjustments led to modest improvements in explained variance (R²), with mean gains of 0.0623% in GEM-pRF and 0.1220% in mrVista. These findings suggest that improving the accuracy of pRF mapping results using refine-fitting might not be necessary if the grid is dense and meets the accuracy requirements of the given application.

Acknowledgements: This research was funded in whole or in part by the Austrian Science Fund (FWF) [https://www.doi.org/10.55776/P35583] and [https://www.doi.org/10.55776/PAT8722623]