Tractometry-based Harmonization of Scanner Effects for White Matter Tract Analysis
Poster No:
1617
Submission Type:
Abstract Submission
Authors:
Bramsh Chandio1, Julio Villalón-Reina1, Talia Nir2, Sophia Thomopoulos3, Yixue Feng4, Neda Jahanshad5, Jaroslaw Harezlak6, Eleftherios Garyfallidis7, Paul Thompson8
Institutions:
1University of Southern California, Los Angeles, CA, 2University of Southern California Keck School of Medicine, Marina del Rey, CA, 3USC, Marina del Rey, CA, 4University of Southern California, Marina Del Rey, CA, 5USC Keck School of Medicine, Marina Del Rey, CA, 6Indiana University School of Public Health, Bloomington, IN, 7Indiana University Bloomington, Bloomington, IN, 8Imaging Genetics Center, Keck School of Medicine of University of Southern California, Los Angeles, CA
First Author:
Co-Author(s):
Paul Thompson, PhD
Imaging Genetics Center, Keck School of Medicine of University of Southern California
Los Angeles, CA
Imaging Genetics Center, Keck School of Medicine of University of Southern California
Los Angeles, CA
Introduction:
In neuroimaging research, the harmonization of data acquired from different imaging sites and scanners is paramount to ensure the reliability, comparability, and accurate interpretation of the findings. ComBat [1,2] is a harmonization method often used to model variations arising from site-specific and scanner-specific effects when analyzing multi-site datasets. It has mostly been applied at the voxel or region of interest level for harmonizing morphometric or microstructural measures when analyzing the measures in group statistics. Here, we add the ComBat method to our BUndle ANalytic (BUAN) [3] tractometry pipeline. We use ADNI3 [4] (Alzheimer's Disease Neuroimaging Initiative) data from 7 diffusion MRI protocols to correct for scanner acquisition protocol effects examining the effects of mild cognitive impairment (MCI) and dementia on 38 white matter tracts. We compare results with and without harmonization using different statistical approaches.
Methods:
Data from 730 ADNI3 participants (age: 55-95, 349M/381F, 214 MCI, 69 dementia, 447 cognitively healthy controls (CN)) scanned with 7 acquisition protocols (GE36, GE54, P33, P36, S127, S31, S55) were included. Tables 1 and 2 detail demographic and acquisition protocol information.
Data was preprocessed using the ADNI3 protocol [5,6], robust and unbiased model-based spherical deconvolution [7], with particle filtering tractography [8] applied to generate tractograms using DIPY [9]. We used BUAN to extract 38 bundles and create their bundle profiles for 4 microstructural metrics, FA, MD, AD, and RD (see Fig. 1 for full names). Bundle profiles consist of 100 segments along the length of the tract with microstructural metrics associated with them. The average of each segment's points is calculated resulting in 100 microstructural values per bundle profile. For each tract and metric, we first pool together bundle profiles of all subjects across CN, MCI, and dementia groups. Pooled bundle profiles consisting of 100 segments modeled as features are then fed to ComBat for scanner protocol effect correction with group, age, and sex modeled as covariates. Due to limited subjects scanned with specific protocols at some sites in ADNI3, we chose not to model site nested within protocol, as the modeling would not be robust. ComBat corrected data is then used to run a linear regression with group, age, and sex modeled as fixed effects and the response variable being microstructural measure, and protocol is incorporated as a random effect in linear mixed models (LMM) to find group differences between MCI and CN and dementia and CN.
Data was preprocessed using the ADNI3 protocol [5,6], robust and unbiased model-based spherical deconvolution [7], with particle filtering tractography [8] applied to generate tractograms using DIPY [9]. We used BUAN to extract 38 bundles and create their bundle profiles for 4 microstructural metrics, FA, MD, AD, and RD (see Fig. 1 for full names). Bundle profiles consist of 100 segments along the length of the tract with microstructural metrics associated with them. The average of each segment's points is calculated resulting in 100 microstructural values per bundle profile. For each tract and metric, we first pool together bundle profiles of all subjects across CN, MCI, and dementia groups. Pooled bundle profiles consisting of 100 segments modeled as features are then fed to ComBat for scanner protocol effect correction with group, age, and sex modeled as covariates. Due to limited subjects scanned with specific protocols at some sites in ADNI3, we chose not to model site nested within protocol, as the modeling would not be robust. ComBat corrected data is then used to run a linear regression with group, age, and sex modeled as fixed effects and the response variable being microstructural measure, and protocol is incorporated as a random effect in linear mixed models (LMM) to find group differences between MCI and CN and dementia and CN.
Results:
After ComBat harmonization, group differences were more sensitively detected, as compared to linear regression and LMM applied without scanner correction. However, we also find comparable results when using LMM with scanner protocol modeled as a random effect aligned with our previous work [4, 5]. While harmonization enhances the significance of the group effect for most tracts, it also decreases the group difference effect for some tracts, suggesting the removal of site-specific confounds that are correlated with the disease effect. We also visualize the results of harmonized BUAN mapping effects of MCI and Dementia, for the left cingulum and right arcuate fasciculus (Fig. 1, bottom panel). Adding different scanner protocols in the analysis boosts disease sensitivity, and harmonization further boosts power, as shown before and after harmonization in Fig. 2f-g.
Conclusions:
Integrating ComBat into the BUAN tractometry can merge data from different scanning protocols. By incorporating ComBat harmonization, we model site-specific and scanner-specific effects, ensuring the reliability and comparability of results by mitigating confounding variables. To the best of our knowledge, this is the first time harmonization has been applied to tractometry analysis. Future work will explore varied ComBat versions and advanced deep-learning approaches.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 2
Modeling and Analysis Methods:
Diffusion MRI Modeling and Analysis 1
Exploratory Modeling and Artifact Removal
Methods Development
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
White Matter Anatomy, Fiber Pathways and Connectivity
Keywords:
Degenerative Disease
Open-Source Software
Statistical Methods
Tractography
White Matter
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
Other - Tractometry, Data Harmonization, ComBat
1|2Indicates the priority used for review
Provide references using author date format
[2] Fortin, J.P., Cullen, N., Sheline, Y.I., Taylor, W.D., Aselcioglu, I., Cook, P.A., Adams, P., Cooper, C., Fava, M., McGrath, P.J. & McInnis, M., 2018. Harmonization of cortical thickness measurements across scanners and sites. Neuroimage, 167, pp.104-120.
[3] Chandio, B.Q., Risacher, S.L., Pestilli, F., Bullock, D., Yeh, F.C., Koudoro, S., Rokem, A., Harezlak, J. & Garyfallidis, E., 2020. Bundle analytics, a computational framework for investigating the shapes and profiles of brain pathways across populations. Scientific Reports, 10(1), pp.1-18.
[4] Zavaliangos-Petropulu, A., Nir, T.M., Thomopoulos, S.I., Reid, R.I., Bernstein, M.A., Borowski, B., Jack Jr, C.R., Weiner, M.W., Jahanshad, N. & Thompson, P.M., 2019. Diffusion MRI indices and their relation to cognitive impairment in brain aging: the updated multi-protocol approach in ADNI3. Frontiers in Neuroinformatics, 13, p.2.
[5] Thomopoulos, S.I., Nir, T.M., Villalon-Reina, J.E., Zavaliangos-Petropulu, A., Maiti, P., Zheng, H., Nourollahimoghadam, E., Jahanshad, N., Thompson, P.M., for the Alzheimer’s Disease Neuroimaging Initiative. Diffusion MRI Metrics and their Relation to Alzheimer’s Disease Severity: Effects of Harmonization Approaches. SIPAIM 2021, Campinas, Brazil. 2021.
[6] Jahanshad, N., Kochunov, P.V., Sprooten, E., Mandl, R.C., Nichols, T.E., Almasy, L., Blangero, J., Brouwer, R.M., Curran, J.E., de Zubicaray, G.I. & Duggirala, R., 2013. Multi-site genetic analysis of diffusion images and voxelwise heritability analysis: A pilot project of the ENIGMA–DTI working group. Neuroimage, 81, pp.455-469.
[7] Canales-Rodríguez, E.J., Daducci, A., Sotiropoulos, S.N., Caruyer, E., Aja-Fernández, S., Radua, J., Yurramendi Mendizabal, J.M., Iturria-Medina, Y., Melie-García, L., Alemán-Gómez, Y. & Thiran, J.P., 2015. Spherical deconvolution of multichannel diffusion MRI data with non-Gaussian noise models and spatial regularization. PloS one, 10(10), p.e0138910.
[8] Girard, G., Whittingstall, K., Deriche, R., & Descoteaux, M. Towards quantitative connectivity analysis: reducing tractography biases. NeuroImage, 98, 266-278, 2014.
[9] Garyfallidis, E., Brett, M., Amirbekian, B., Rokem, A., van der Walt, S., Descoteaux, M. & Nimmo-Smith, I., “Dipy, a library for the analysis of diffusion MRI data,” Front. Neuroinform. 8 8 (2014).