Comprehensive evaluation of harmonization on functional brain imaging for multisite data-fusion

Abstract

To embrace big-data neuroimaging, harmonization of site effect in resting-state functional magnetic resonance imaging (R-fMRI) data fusion is a fundamental challenge. Comprehensive evaluation of potentially effective harmonization strategies, particularly with specifically collected data has been rare, especially for R-fMRI metrics. Here, we comprehensively assess harmonization strategies from multiple perspectives, including efficiency, individual identification, test-retest reliability and replicability of group-level statistical results, on widely used R-fMRI metrics across multiple datasets including data obtained from the same participants scanned at several sites. For individual identifiability (i.e., whether the same subject could be identified across R-fMRI data scanned across different sites), we found that, while most methods decreased site effects, the Subsampling Maximum-mean-distance based distribution shift correction Algorithm (SMA) outperformed linear regression models, linear mixed models, ComBat series and invariant conditional variational auto-encoder. Test-retest reliability was better for SMA and adjusted ComBat series than alternatives, while SMA was superior to the latter in replicability, both in terms of Dice coefficient and the scale of brain areas showing sex differences reproducibly observed across datasets. Moreover, we examined test-retest datasets to identify the best target site features to optimize SMA identifiability and test-retest reliability. We noted that both sample size and distribution of the target site matter and introduced a heuristic target site selection formula. In addition to providing practical guidelines, this work can inform continuing improvements and innovations in harmonizing methodologies for big R-fMRI data.

Brain‐age prediction: Systematic evaluation of site effects, and sample age range and size

“…We evaluated seven site handling strategies after partitioning the discovery sample into different age bins as follows: (i) a single bin with the full sample age range (5–90 years); (ii) nine bins each covering sequential 10‐year intervals, that is, age ≤ 10 years, 10 < age ≤ 20 years, 20 < age ≤ 30 years, 30 < age ≤ 40 years, 40 < age ≤ 50 years, 50 < age ≤ 60 years, 60 < age ≤ 70 years, 70 < age ≤ 80 years, and 80 < age ≤ 90 years; (iii) four bins each covering sequential 20‐year intervals, that is, age ≤ 20 years, 20 < age ≤ 40 years, 40 < age ≤ 60 years, and 60 < age ≤ 80 years; (iv) three bins each covering sequential 30‐year intervals, that is, age ≤ 30 years, 30 < age ≤ 60 years, and 60 < age ≤ 90 years; (v) two age bins each covering sequential 40‐year intervals, that is, age ≤ 40 years and 40 < age ≤ 90 years. Seven site handling strategies were separately applied to each bin to perform data residualization with respect to site using: (i) Combat‐GAM (Pomponio et al, 2020); (ii) CovBat without age variability preservation (Chen et al, 2021); (iii) CovBat with age variability preservation (Chen et al, 2021); (iv) Subsampling Maximum‐mean‐distance Algorithm (SMA) (Wang et al, 2023; Zhou et al, 2018); (v) Invariant Conditional Variational Auto‐Encoder (ICVAE) (Moyer et al, 2020; Wang et al, 2023); (vi) generalized linear model (de Lange et al, 2022); and (vii) no site harmonization. In the case of Combat‐GAM, age was specified as the smooth term in the model while the empirical Bayes estimates were used for site effects, without custom boundaries for the smoothing terms.…”

“…The CovBat approach was implemented using R scripts (version 3.6.0). The SMA method was implemented using Matlab (version R2021a) with the largest sample as the target site, in accordance with recommendations of Wang et al (2023) The ICVAE was implemented using Python (version 3.8.10). To prevent data leakage, the harmonization process was applied separately to the training and test datasets during cross‐validation.…”

Low Functional network integrity in cognitively unimpaired and MCI subjects with depressive symptoms: results from a multi-center fMRI study

“…The interaction of depression and MCI was included in all GLM analyses in order to study the possible different effects of depressive symptoms on network functionality in MCI and cognitively healthy subjects. Furthermore, the analyses included MRI scanner type [60], sex, age, education, and total connectivity strength as covariates. The same model was applied to analyze differences in network metrics and structural measures (e.g., cortical thickness and volumes).…”

Tractometry of Human Visual White Matter Pathways in Health and Disease

“…This challenge is not unique to dMRI data and affects other MRI quantities such as cortical thickness assessed from T1-weighted and T2-weighted imaging, 236 as well as statistics derived from fMRI measurements. 237 , 238 Significant research effort has been invested in developing methods for signal harmonization across different scanners 239 – 241 but the state of the art in these methods still requires measurements from many different subjects in each scanner. 242 , 243 This makes the clinical application of dMRI tractometry challenging because comparing an individual’s brain against a normative sample would have to happen either on the same scanner as the one on which the normative sample was acquired or on a scanner that has been harmonized with the scanner used for normative measurements (there are some important exceptions to this rule; for example, where a disease affects lateralization of the measurement 244 ).…”