Optimizing FreeSurfer's Surface Reconstruction Parameters for Anatomical Feature Estimation

Magnetic resonance imaging (MRI) is a powerful tool for non-invasive imaging of the human body. However, the quality and reliability of MRI data can be influenced by various factors, such as hardware and software configurations, image acquisition protocols, and preprocessing techniques. In recent years, the introduction of large-scale neuroimaging datasets has taken an increasingly prominent role in neuroscientific research. The advent of publicly available and standardized repositories has enabled researchers to combine data from multiple sources to explore a wide range of scientific inquiries. This increase in scale allows the study of phenomena with smaller effect sizes over a more diverse sample and with greater statistical power. Other than the variability inherent to the acquisition of the data across sites, preprocessing and feature generation steps implemented in different labs introduce an additional layer of variability which may influence consecutive statistical procedures. In this study, we show that differences in the configuration of surface reconstruction from anatomical MRI using FreeSurfer results in considerable changes to the estimated anatomical features. In addition, we demonstrate the effect these differences have on within-subject similarity and the performance of basic prediction tasks based on the derived anatomical features. Our results show that although FreeSurfer may be provided with either a T2w or a FLAIR scan for the same purpose of improving pial surface estimation (relative to based on the mandatory T1w scan alone), the two configurations have a distinctly different effect. In addition, our findings indicate that the similarity of within-subject scans and performance of a range of models for the prediction of sex and age are significantly effected, they are not significantly improved by either of the enhanced configurations. These results demonstrate the large extent to which elementary and sparsely reported differences in preprocessing workflow configurations influence the derived brain features. The results of this study are meant to underline the importance of optimizing preprocessing procedures based on experimental results prior to their distribution and consecutive standardization and harmonization efforts across public datasets. In addition, preprocessing configurations should be carefully reported and included in any following analytical workflows, to account for any variation originating from such differences. Finally, other representations of the raw data should be explored and studied to provide a more robust framework for data aggregation and sharing.