Siddharth Mittal¹ (), Michael Woletz¹, David Linhardt¹, Christian Windischberger¹; ¹Medical University of Vienna

Population-receptive field (pRF) mapping is a tool for mapping visual information encoding in the brain using fMRI. The pRFs are typically modelled as 2D Gaussian, where the mean (μ_x, μ_y) signifies location, and the variance (σ) indicates the receptive field size. Traditional pRF mapping tools like mrVista and SamSrf yield accurate results but require long computation times. Conversely, newer methods like fast-pRF from CNI_toolbox favour speed over accuracy. To bridge this gap, we propose the novel implementation GEM-pRF (GPU-Empowered Mapping of pRF), which combines high accuracy with greatly shortened computation time. This method involves two steps: (1) initial grid search for pRF positions and size estimates (μ_x, μ_y, σ) executed through efficient matrix operations on high-performance GPU cores and (2) approximating the residual sum of squares (RSS) error function as quadratic, using partial derivatives in the neighbourhood to refine estimates. We evaluated the accuracy of our results using simulated and real fMRI data. For simulated data, we employed a validation framework by Lerma-Usabiaga et al. (2020), to create a noisy simulated dataset with known pRF parameters. Our implementation (GEM-pRF) was compared with mrVista and SamSrf, showing similarly high accuracy in pRF parameter estimation. For analysis on real data, we scanned a healthy male on a 3T Siemens PrismaFit scanner (CMRR EPI, TR/TE=1000/38ms, 1.5mm iso, MB=3). Using this real data, we compared our method's pRF parameter estimations with mrVista (a commonly used tool for pRF mapping). The results revealed a high correspondence between the two. Notably, our method estimates pRF parameters for a 10,000-voxel fMRI dataset in just 30-40 seconds, a significant improvement from the approx. 10 minutes taken by mrVista. These findings underscore the remarkable speed and maintained accuracy of our GPU-based implementation, enabling comprehensive analysis of large datasets and unlocking new possibilities for exploration with complex pRF models.

Acknowledgements: This work was supported by the Austrian Science Fund (FWF; grant numbers: P35583)