GPU-Empowered Mapping of Population Receptive Fields (GEM-pRF)
Poster No:
1349
Submission Type:
Abstract Submission
Authors:
Siddharth Mittal1, Michael Woletz1, David Linhardt2, Christian Windischberger3
Institutions:
1Medical University of Vienna, Vienna, Vienna, 2Medical University of Vienna, Wien, Vienna, 3Medical University of Vienna, Vienna, Austria
First Author:
Co-Author(s):
Introduction:
Population receptive field (pRF) mapping allows for determining the retinotopic organisation in the visual cortex (Dumoulin & Wandell, 2008). It involves modelling pRFs as 2D Gaussian functions in the visual field and fitting them to time series data of the neuronal responses. While highly accurate, this process is computationally demanding due to iterative refinement.
To address this, we propose a novel pRF implementation called GPU-Empowered Mapping of pRF (GEM-pRF). It uses a reformulated mathematical notation based on the originally proposed technique by Dumoulin & Wandell (2008) for faster and accurate pRF parameter estimation. Here we show that our new approach harnesses high-performance computing advantages to speed up acquisition by over 15 times.
To address this, we propose a novel pRF implementation called GPU-Empowered Mapping of pRF (GEM-pRF). It uses a reformulated mathematical notation based on the originally proposed technique by Dumoulin & Wandell (2008) for faster and accurate pRF parameter estimation. Here we show that our new approach harnesses high-performance computing advantages to speed up acquisition by over 15 times.
Methods:
GEM-pRF is a two-stage procedure. During the first stage, a grid search is performed to compute the coarse pRF parameters estimation. The estimated pRF parameters for a given pRF are its position and size (i.e., μx, μy, and σ). Our derived mathematical notation simplifies this step in order to use GPU-based matrix multiplication for major speed gains. In the second stage, refining the coarsely estimated pRF parameters involves a quadratic approximation of the error function in the neighbourhood of the coarse result. We use partial derivatives of the error term to find minima, thereby enabling estimation of the best pRF fitting parameters. Goodness-of-fit is based on residual sum of squares (RSS). Our novel approach was evaluated on a laptop with an NVIDIA RTX 3050Ti GPU and a university server's GPU cluster using NVIDIA Tesla V100. We assessed the accuracy of this method with both simulated and empirical data.
Results:
Our GEM-pRF implementation shows notable speed enhancement without compromising accuracy. To validate our results on simulated data, we utilised a validation framework (Lerma-Usabiaga et al., 2020) for pRF mapping methods. Within this framework, we generated low- and high-noise datasets comprising 5000 voxels each. Datasets were generated for two scenarios, first for a smaller pRF located at visual field centre (i.e. μx=0, μy=0, σ=1), and second for off-centred bigger pRF (i.e. μx=3, μy=-2 , σ=1.8). We compared our pRF parameter estimations with different fMRI analysis toolboxes (mrVista , SamSrf, and f-pRF ((Bhat et al., 2021) from CNI_toolbox). In Figure 1, our new implementation pRF parameter estimation (position & size) showed very good pRF parameter estimation results compared to other pRF implementations in both low- and high SNR scenarios.
For evaluation on empirical data, we scanned a healthy male using a 64-channel head coil on a 3T Siemens PrismaFit scanner. The full coil was used for anatomical measurements (MP2RAGE, 1mm iso), while functional measurements (CMRR EPI, TR/TE=1000/38ms, 1.5mm iso, MB=3) used the coil's lower part only. Our pRF parameter estimates were compared to those computed by the standard mrVista implementation, a specialised software package for pRF fMRI data analysis. The comparison of results is presented in Figure 2, which shows the estimated pRF positions (μx, μy) and their sizes (σ). Figure 2(d) shows the comparison of Variance Explained (R2) values for the modelled signals. The overall comparison reveals high correspondence between our novel approach and the standard mrVista implementation. Our implementation, however, requires considerably less computation time: while mrVista and similar implementations require about 10 minutes, our implementation completes pRF parameter estimation in approx. 30-40 seconds for datasets containing up to 10,000 voxels.
For evaluation on empirical data, we scanned a healthy male using a 64-channel head coil on a 3T Siemens PrismaFit scanner. The full coil was used for anatomical measurements (MP2RAGE, 1mm iso), while functional measurements (CMRR EPI, TR/TE=1000/38ms, 1.5mm iso, MB=3) used the coil's lower part only. Our pRF parameter estimates were compared to those computed by the standard mrVista implementation, a specialised software package for pRF fMRI data analysis. The comparison of results is presented in Figure 2, which shows the estimated pRF positions (μx, μy) and their sizes (σ). Figure 2(d) shows the comparison of Variance Explained (R2) values for the modelled signals. The overall comparison reveals high correspondence between our novel approach and the standard mrVista implementation. Our implementation, however, requires considerably less computation time: while mrVista and similar implementations require about 10 minutes, our implementation completes pRF parameter estimation in approx. 30-40 seconds for datasets containing up to 10,000 voxels.
Conclusions:
Our proposed GPU-empowered mapping of pRF (GEM-pRF) approach maintains the high-accuracy of the pRF parameters estimation results as compared to the existing popular techniques, while reducing the computational time by more than an order of magnitude.
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI) 1
Novel Imaging Acquisition Methods:
BOLD fMRI 2
Keywords:
Computational Neuroscience
FUNCTIONAL MRI
HIGH FIELD MR
Modeling
Open-Source Code
Open-Source Software
Statistical Methods
Vision
Other - Retinotopy, Population Receptive Fields (pRF) Mapping
1|2Indicates the priority used for review
Provide references using author date format
Lerma-Usabiaga, G., Benson, N., Winawer, J., & Wandell, B. A. (2020). A validation framework for neuroimaging software: The case of population receptive fields. PLoS Computational Biology, 16(6). https://doi.org/10.1371/journal.pcbi.1007924
Bhat, S., Lührs, M., Goebel, R., & Senden, M. (2021). Extremely fast pRF mapping for real-time applications. NeuroImage, 245. https://doi.org/10.1016/j.neuroimage.2021.118671