Optimizing the preprocessing pipeline for multicenter fMRI studies

1956 

Abstract Submission 

Marie-Eve Hoeppli¹, Emma Biggs², Saül Pascual-Diaz³, Massieh Moayedi⁴, Marina López-Solà⁵, Laura Simons⁶, Robert Coghill⁷

¹CCHMC, Cincinnati, OH, ²Department of Anesthesiology, Perioperative and Pain Medicine, Stanford University School of Medici, Stanford, CA, ³Universitat de Barcelona, Barcelona, Spain, ⁴Centre for Multimodal Sensorimotor and Pain Research, Faculty of Dentistry, University of Toronto, Toronto, Ontario, ⁵University of Barcelona, Barcelona, Barcelona, ⁶Department of Anesthesiology, Perioperative and Pain Medicine, Stanford University School of Medicin, Stanford, CA, ⁷Cincinnati Children's Hospital Medical Center, Cincinnati, OH

Marie-Eve Hoeppli  
CCHMC
Cincinnati, OH

Emma Biggs  
Department of Anesthesiology, Perioperative and Pain Medicine, Stanford University School of Medici
Stanford, CA

Saül Pascual-Diaz  
Universitat de Barcelona
Barcelona, Spain

Massieh Moayedi  
Centre for Multimodal Sensorimotor and Pain Research, Faculty of Dentistry, University of Toronto
Toronto, Ontario

Marina López-Solà  
University of Barcelona
Barcelona, Barcelona

Laura Simons  
Department of Anesthesiology, Perioperative and Pain Medicine, Stanford University School of Medicin
Stanford, CA

Robert Coghill  
Cincinnati Children's Hospital Medical Center
Cincinnati, OH

Today an increasing number of fMRI analyses relies on advanced statistical techniques based on machine learning approaches. These techniques require large sample sizes, which can be challenging to achieve in specific population, e.g. children with chronic musculoskeletal pain. Combining the recruitment efforts of multiple research centers is then needed to successfully complete the required sample size. Due to differences in equipment and practices, this multicenter approach can render the optimization of the protocols and the processing of the data complicated. In this study we aimed at defining the optimal preprocessing pipeline for data acquired in the context of a large-scale study on three different fMRI 3T scanners: Siemens Prisma, GE Premier, and Philips Ingenia Elition (Simons et al., 2022).

Of 131 teenagers diagnosed with chronic musculoskeletal pain who were enrolled in the main study and underwent an fMRI session, 30 were selected for this substudy. 10 participants from each site were included. Participants were selected on the noise observed in the raw data to ensure an equal representation of noisy and less noisy data. Each participant underwent two resting-state multiband BOLD sequences. Sequence parameters, including reaction time, echo time, flip angle, and voxel size, were constant across centers (Simons et al., 2022).
Three main preprocessing pipelines were evaluated: CONN (Whitfield-Gabrieli & Nieto-Castanon, 2012), fMRIprep, and FSL (Jenkinson, Beckmann, Behrens, Woolrich, & Smith, 2012). In addition, three denoising techniques, ICA-AROMA (Pruim et al., 2015), FSL FIX with custom-trained classifier (Salimi-Khorshidi et al., 2014), and aCompCor (Behzadi, Restom, Liau, & Liu, 2007) were tested in a standard FSL pipeline. Finally, a comparison of preprocessing pipelines including temporal high-pass filtering (cut-off: 125sec.) was compared pipelines including temporal bandpass filtering (cut-offs: 125sec. and 11sec.). Signal-to-noise ratios were calculated for each participant and pipeline and compared between pipelines within predefined brain areas, including amygdala, insula, cingulate cortex, ventromedial prefrontal cortex, orbitofrontal cortex, primary and secondary somatosensory cortices, hippocampus, thalamus, cerebrospinal fluid, and white matter.

Results showed that (1) across the three main pipelines (CONN, FSL, and fMRIprep), applying bandpass filtering improved the remaining signal; (2) FSL performed better than fMRIprep or CONN; (3) the application of ICA-based denoising techniques in FSL, and in particular FSL FIX, further improved signal conservation.

In conclusions, our results suggest that an FSL-based preprocessing pipeline including FIX as noise reduction technique and bandpass filtering might be the optimal preprocessing pipeline for multicenter approaches. These results are in line with previous results (Hoeppli et al., 2023), highlighting the benefit of FSL FIX on the quality of preprocessed data. Furthermore, these results demonstrate the advantage of applying bandpass filtering on resting-state BOLD data. In summary, the preprocessing of fMRI data can be particularly challenging in studies relying on acquisitions from different centers and requires careful consideration. Our results are a first attempt to provide guidance to optimize this process.

Early life, Adolescence, Aging

Motion Correction and Preprocessing ¹

BOLD fMRI ²

Perception: Pain and Visceral

FUNCTIONAL MRI

Pain

PEDIATRIC