3002
ComBat harmonization for multi-site fixel-based analysis using traveling subject dataset1Department of Data Science, Juntendo University Graduate School of Medicine, Tokyo, Japan, 2Department of Radiology, Juntendo University Graduate School of Medicine, Tokyo, Japan, 3Faculty of Health Data Science, Juntendo University, Chiba, Japan, 4Department of Radiological Sciences, Graduate School of Human Health Sciences, Tokyo Metropolitan University, Tokyo, Japan, 5Department of Radiology, The University of Tokyo, Tokyo, Japan
Synopsis
Keywords: Data Processing, Data Processing, Diffusion MRI, harmonization, fixel-based analysis, multisite
Motivation: Although multi-site DWI with large sample size has high statistical power and is sensitive to the subtle microstructural tissue changes, different models or protocols-induced measurement biases affect the reliability and reproducibility of the study. Therefore, harmonization is necessary to improve this issue.
Goal(s): The goal of our study is to evaluate the effectiveness of ComBat harmonization in mitigating measurement biases in FBA measures.
Approach: Our study utilized a traveling-subject DWI dataset, while various FBA measures were calculated and subsequently harmonized using the ComBat method.
Results: Our findings demonstrated that ComBat harmonization could effectively mitigate site, model, and protocol-induced measurement biases in FBA measures.
Impact: A significant contribution of this study is the seamless integration of ComBat into the fixel-based framework, which may enhance the reliability and reproducibility of multi-site research, offering a valuable tool for investigating microstructural tissue changes in the large-scale, multi-site studies.
Introduction
Multi-site diffusion-weighted magnetic resonance imaging (DWI) with large sample size is increasingly used in clinical research and development1-3, and the measures derived from multi-site DWI have high statistical power and are quite sensitive to the subtle changes of the underlying microstructural tissue, contributing to a wealth of knowledge about the abnormalities in several neurological and psychiatric disorders4-5. In advanced DWI techniques, fixel-based analysis (FBA) is becoming a popular approach that can provide fiber-specific estimates of white matter abnormality6-7 and be expected to use for multi-site DWI study. However, in the multi-site DWI data, there are measurement biases-caused by site, model, and protocol differences-that influence the reliability and reproducibility of DWI analysis8-9. Therefore, harmonization method, which can remove the measurement bias, is necessary to improve the reproducibility for the multi-site DWI study. Herein, we implemented the combined association test (ComBat) method, a popular harmonization approach for neuroimaging data10-11, on the FBA measures derived from a multi-site DWI dataset including traveling subjects scanned with different models and protocols in different sites.Methods
We acquired a DWI traveling-subject dataset including 162 scans of 49 healthy participants across 6 different research sites, employing 3 Siemens models of a single 3T MRI vendor and 2 distinct scanning protocols12. To investigate the effect of measurement bias, DWI scans were divided into four groups including scan-rescan, site-, model-, and protocol-difference as depicted in Table 1. Each MRI acquisition utilized a 32-channel head coil and implemented a single-shot spin-echo-planar sequence with monopolar diffusion gradients. The specific MRI acquisition parameters for each scanning protocol can be found in Table 2.We performed preprocessing of the DWI data using MRtrix313. Fiber Orientation Distributions were computed through multi-shell 3-tissue constrained spherical convolution14, and spatial alignment was achieved by generating a population template image based on a randomly selected cohort of 40 individuals. Measures of fiber density (FD), log-transformed fiber bundle cross- section (logFC), and fiber density and cross-section (FDC) were calculated for each white matter fixel and subsequently smoothed in fixel-based connectivity analysis.To harmonize measurement bias, we applied the ComBat to the FBA measures. Subsequently, we assessed the statistically significant differences (p < 0.05) between site-, model-, and protocol-difference in both before and after ComBat harmonization.
Results
Before harmonization, there is no differences of FD, logFC and FDC for scan-rescan and slight differences in partial area for site difference. However, we observed significant differences in FD, logFC and FDC for model and protocol differences. Notably, the relative error was approximately 8-15%. After ComBat harmonization, all significant differences of FD, logFC and FDC between different sites were eliminated. The differences between different models and protocols were much smaller than which before harmonization. Regarding model differences, there remained a slight relative error of 2-3% in the corpus callosum and the frontal white matter. In terms of protocol differences, a relative error of about 2-3% persisted in the cerebellum (Figure 1).Discussion
In previous studies, ComBat is useful for mitigating the measurement bias in diffusion tensor imaging (DTI)11. Besides, Mito R et al. demonstrated that ComBat could remove measurement bias between different models in FDC of FBA measures15. Furthermore, by using the traveling subject data which includes only measurement bias, but not biological sampling bias (i.e., differences between the participant groups), we purely indicated the effective capacity of ComBat in mitigating measurement bias from site, model, and protocol difference within the DWI dataset.In previous FBA investigations of neurodegenerative and neuropsychiatric disorders, the differences of FD, FC, FDC values between healthy subjects and patients was approximately 10-20% for Alzheimer's disease16, roughly 5-10% in Parkinson's disease17-18, and about 3-10% in major depressive disorder and chronic schizophrenia19-20. These differences between healthy subjects and patients are comparable to, or even less than the measurement biases (8-15%) caused by different models and protocols before harmonization, indicating that the diseases might be overlooked because the differences caused by diseases are indistinguishable from the measurement biases. Interestingly, following the application of ComBat harmonization in our study, we observed a notable reduction of FBA value variances to approximately 2-3%. This outcome suggests that harmonizing FBA measures using ComBat has the potential to enhance the sensitivity for detecting disease-induced variations.In conclusion, our findings demonstrated that ComBat harmonization could effectively mitigate site, model, and protocol-induced measurement bias of FBA measures. These findings hold the potential of utilizing harmonization to improve the reliability and reproducibility of multi-site DWI study, anticipating a transition from single-site studies with limited sample sizes to large-scale, multi-site DWI investigations with high statistical power in the future.
Acknowledgements
This research was supported by Brain/MINDS Beyond program (grant no. JP18dm0307006, JP18dm0307001, JP18dm0307004, JP18dm0307008, JP19dm0207069, JP19dm0307101, and JP22dm0307002) of the Japan Agency for Medical Research and Development (AMED), AMED under grant number JP21wm0425006, JSPS KAKENHI (grant nos. 20K16737, 21K07690, 21K12153, 21K15833, 22H04926, 23H02865), a Grant-in-Aid for Special Research in Subsidies for ordinary expenses of private schools from The Promotion and Mutual Aid Corporation for Private Schools of Japan, and the Juntendo Research Branding Project. This study was also supported by the World Premier International-International Research Center for Neurointelligence (WPI-IRCN) and Japan Science and Technology Agency (JST) Moonshot R&D Grant Number JPMJMS2021.This work was also supported by the Otsuka Toshimi Scholarship Foundation (to S. Guo.).References
1. Sprague S, Matta JM, Bhandari M, et al. Multicenter collaboration in observational research: improving generalizability and efficiency. J Bone Joint Surg Am. 2009;91.
2. Schilling KG, Archer D, Yeh FC, et al. Aging and white matter microstructure and macrostructure: a longitudinal multi-site diffusion MRI study of 1218 participants. Brain Struct Funct. 2022;227(6):2111-2125.
3. Pohl KM, Sullivan EV, Rohlfing T, et al. Harmonizing DTI measurements across scanners to examine the development of white matter microstructure in 803 adolescents of the NCANDA study. Neuroimage. 2016; 130:194-213.
4. Kubicki M, McCarley R, Westin CF, et al. A review of diffusion tensor imaging studies in schizophrenia. J Psychiatr Res. 2007;41(1-2):15-30.
5. Shenton ME, Hamoda HM, Schneiderman JS, et al. A review of magnetic resonance imaging and diffusion tensor imaging findings in mild traumatic brain injury. Brain Imaging Behav. 2012;6(2):137-192.
6. Raffelt DA, Tournier JD, Smith RE, et al. Investigating white matter fibre density and morphology using fixel-based analysis. Neuroimage. 2017;144(Pt A):58-73.
7. Grazioplene RG, Bearden CE, Subotnik KL, et al. Connectivity-enhanced diffusion analysis reveals white matter density disruptions in first episode and chronic schizophrenia. Neuroimage Clin. 2018; 18:608-616.
8. Yamashita A, Yahata N, Itahashi T, et al. Harmonization of resting-state functional MRI data across multiple imaging sites via the separation of site differences into sampling bias and measurement bias. PLoS Biol. 2019;17(4): e3000042.
9. Andica C, Kamagata K, Hayashi T, et al. Scan-rescan and inter-vendor reproducibility of neurite orientation dispersion and density imaging metrics. Neuroradiology. 2020;62(4):483-494.
10. Beer JC, Tustison NJ, Cook PA, et al. Longitudinal ComBat: A method for harmonizing longitudinal multi-scanner imaging data. Neuroimage. 2020; 220:117129.
11. Fortin JP, Parker D, Tunç B, et al. Harmonization of multi-site diffusion tensor imaging data. Neuroimage. 2017; 161:149-170.
12. Koike S, Tanaka SC, Okada T, et al. Brain/MINDS beyond human brain MRI project: A protocol for multi-level harmonization across brain disorders throughout the lifespan. Neuroimage Clin. 2021; 30:102600.
13. Tournier JD, Smith R, Raffelt D, et al. MRtrix3: A fast, flexible and open software framework for medical image processing and visualisation. Neuroimage. 2019; 202:116137.
14. Jeurissen B, Tournier JD, Dhollander T, Connelly A, Sijbers J. Multi-tissue constrained spherical deconvolution for improved analysis of multi-shell diffusion MRI data. Neuroimage. 2014; 103:411-426.
15. Mito R, Pardoe H, Smith R, et al. ComBat scanner harmonization for fixel-based analysis. The 2023 Annual Meeting of International Society for Magnetic Resonance in Medicine. Toronto, Canada, 2023.
16. Mito R, Raffelt D, Dhollander T, et al. Fibre-specific white matter reductions in Alzheimer's disease and mild cognitive impairment. Brain. 2018;141(3):888-902.
17. Andica C, Kamagata K, Saito Y, et al. Fiber-specific white matter alterations in early-stage tremor-dominant Parkinson's disease. NPJ Parkinsons Dis. 2021;7(1):51.
18. Andica C, Kamagata K, Uchida W, et al. Fiber-Specific White Matter Alterations in Parkinson's Disease Patients with GBA Gene Mutations. Mov Disord. 2023;10.1002.
19. Grazioplene RG, Bearden CE, Subotnik KL, et al. Connectivity-enhanced diffusion analysis reveals white matter density disruptions in first episode and chronic schizophrenia. Neuroimage Clin. 2018; 18:608-616.
20. Lyon M, Welton T, Varda A, et al. Gender-specific structural abnormalities in major depressive disorder revealed by fixel-based analysis. Neuroimage Clin. 2019; 21:101668.
Figures
Table 1. Demographic characteristics and groups of study participants
The DWI data of 49 participants (6 sites, 3 models, a total of 162 scans) was acquired followed by a “hub-and-spoke” design in which participant traveled to 6 sites, but not all sites11. DWI data were assigned into scan–rescan and three measurement bias factors (site, model, and protocol) to evaluate the harmonization performance of ComBat.
Table 2. Acquisition parameters of diffusion MRI
Each DWI examination was performed using a Siemens 3T model with a 32-channel head coil and the presented acquisition parameters. All DWI data were corrected for susceptibility, eddy-current induced geometric distortions, and intervolume subject motion. FBA measures were estimated based on multi-shell 3-tissue constrained spherical convolution using corrected DWI data13.
Figure 1. FBA of DWI data before and after ComBat harmonization
The rows show the relative error of FBA measures (FD, logFC and FDC) with significant difference (p < 0.05) in the whole-brain, while the columns show the sections under scan-rescan and the measurement bias (site, model, and protocol) before and after ComBat harmonization. The color indicates the relative error of the measurement bias in the regional white matter tracts with red indicating a low relative error (relative error of 0%), changing to yellow for a high relative error (relative error of 20%).