DeepResBat: A residual MRI harmonization framework accounting for covariate distribution differences

1848 

Abstract Submission 

Lijun An¹, Chen Zhang¹, Naren Wulan¹, Pansheng Chen¹, Shaoshi Zhang¹, Fang Ji¹, Kwun Kei Ng¹, Christopher Chen¹, Juan Helen Zhou¹, B. T. Thomas Yeo¹

¹National University of Singapore, Singapore, Singapore

Lijun An  
National University of Singapore
Singapore, Singapore

Chen Zhang  
National University of Singapore
Singapore, Singapore

Naren Wulan  
National University of Singapore
Singapore, Singapore

Pansheng Chen  
National University of Singapore
Singapore, Singapore

Shaoshi Zhang  
National University of Singapore
Singapore, Singapore

Fang Ji  
National University of Singapore
Singapore, Singapore

Eric Kwun Kei Ng  
National University of Singapore
Singapore, Singapore

Christopher Chen  
National University of Singapore
Singapore, Singapore

Juan Helen Zhou  
National University of Singapore
Singapore, Singapore

B. T. Thomas Yeo  
National University of Singapore
Singapore, Singapore

Mega-analysis combining MRI data from multiple sites has significantly advanced neuroimaging research (Miller et al., 2018). When pooling data across multiple sites, MRI harmonization is crucial to reduce undesired site variabilities while preserving effects of covariates of interests (e.g., age). Recently, deep learning has been utilized for addressing harmonization problems (Hu et al., 2023). However, existing deep learning harmonization approaches typically ignore the inclusion of covariates (Zuo et al., 2021; Bashyam et al., 2021). Here, we propose deep learning harmonization models accounting for covariate distribution differences. Furthermore, we demonstrate for the first time that deep learning harmonization algorithms can introduce false positives, which should lead to caution for applying deep learning in neuroimaging.

In this study, we harmonized T1 MRI between ADNI (adni.loni.usc.edu) and AIBL (aibl.loni.csiro.au), as well as between ADNI and MACC (macc.sg) cohorts. The inputs obtained from FreeSurfer (Fischl et al., 2002; Desikan et al., 2006) are thickness and volumes of cortical ROIs, and volumes of subcortical ROIs. We performed a matching procedure to control for demographics, disease severity, and cognition impairment differences across cohorts. The matched participants served as the testset. The unmatched participants served as training and validation sets.
We proposed two deep learning models, coVAE (Figure 1B) and DeepResBat (Figure 1C), which account for covariate distribution differences across datasets. coVAE extends cVAE (Figure 1A; Moyer et al., 2020) by concatenating covariates y and site s with latent representations z. DeepResBat preserves the effects of covariates using machine learning and then eliminates unwanted site differences in covariate-free residuals through a cVAE model. For the baselines, we considered ComBat (Johnson et al., 2007), CovBat (Chen et al., 2021), and cVAE.
We considered three evaluation metrics. Firstly, an XGBoost model was applied to predict which dataset a harmonized participant came from. A higher accuracy indicates worse harmonization. Secondly, we tested whether harmonization leads to stronger association between harmonized ROIs and covariates of interests. A stronger association means better harmonization. Finally, we randomly permuted covariates and then harmonized ROIs. The permutation was repeated 1000 times, costing 240,000 GPU and 360,000 CPU hours. For each ROI, a GLM was fitted for each permutation to check association between harmonized ROI and permutated covariates, resulting in 1000 p values; then, we calculated percentage of p values below 0.05 across 1000 p values. The expected percentage across all ROIs is 5%, with a confidence interval of 3.65% to 6.35% (Eklund et al., 2016). A percentage higher than 5% underscores false positives.

Figure 2A shows that deep learning approaches significantly removed more dataset differences than linear approaches (ComBat and CovBat) for harmonizing ADNI and AIBL (Figure 2A1) and harmonizing ADNI and MACC (Figure 2A2) datasets.
Figure 2B shows association analysis results between harmonized ROIs and covariates of interests, measured by multivariate ANOVA. We found that coVAE and DeepResBat preserved more biological variabilities than cVAE without considering covariates and linear approaches.
Our permutation tests (Figure 2C) show coVAE introduced high false positives, while the DeepResBat achieved expected false positives. The results indicated deep learning harmonization approaches can introduce false positives without careful modeling.

In this study, we proposed deep learning models, coVAE and ResBat, for incorporating covariates during harmonization. coVAE and ResBat outperformed cVAE and linear approaches in removing dataset differences and preserving biological variability. Furthermore, DeepResBat achieved acceptable false positives, which is a novel finding for deep learning harmonization studies.

Methods Development ¹

Multivariate Approaches ²

Machine Learning

Modeling

MRI

Workflows

Other - MRI harmonization