Abstract
Population receptive field (pRF) mapping allows for the investigation of retinotopic organization preserved from the retina to the visual cortex. From fMRI acquisition to the final results, the data goes through many transformations and computational methods. Usually, these steps are individually implemented and/or require specific manual interventions and are processed in lab-specific pipelines. Re-running these scripts in different environments often yields variations of fMRI results due to different software versions. In our experience, the combination of dynamic software environments and the sheer number of configuration options thus reduces the reproducibility of pRF analysis results. In this work, we aimed at providing a comprehensive framework for simple, reproducible pRF analysis using a containerized solution. Container technology is the most sophisticated method for code execution. It does not depend on the host operating system and local hard- or software specifications. The container hosts all code snippets, programs and libraries necessary for the execution of the analyses, guaranteeing computational reproducibility and also simplifying complicated workflows. In our solution, we integrate three existing and two new containers: (1) HeuDiConv (github.com/nipy/heudiconv) for DICOM to NIfTI conversion; (2) fMRIPrep (fmriprep.org) for fMRI data pre-processing; (3) the newly developed prfprepare (github.com/dlinhardt/prfprepare) for performing several pRF-specific preparation steps to enable seamless usage of the preprocessed data, (4) an amended version of prfanalyze (GitHub.com/vistalab/PRFmodel) for pRF-analysis; and (5) the novel prfresult (github.com/dlinhardt/prfresult) for visualization of the results. We used this framework on data from 30 subjects to systematically and reproducibly test (1) the effect of using different seeds / initial conditions in 3 different pRF analysis implementations, and (2) the differences in visual field coverage when analyzing fMRI time series represented in volumetric or surface space. We illustrate the results of our analyses using prfresult and provide a set of recommendations for robust and reproducible pRF mapping.