Diffusion MRI harmonization enables joint-analysis of multicentre data of patients with cerebral small vessel disease

Evaluation of Mean Shift, ComBat, and CycleGAN for Harmonizing Brain Connectivity Matrices Across Sites

Connectivity matrices derived from diffusion MRI (dMRI) provide an interpretable and generalizable way of understanding the human brain connectome. However, dMRI suffers from inter-site and between-scanner variation, which impedes analysis across datasets to improve robustness and reproducibility of results. To evaluate different harmonization approaches on connectivity matrices, we compared graph measures derived from these matrices before and after applying three harmonization techniques: mean shift, ComBat, and CycleGAN. The sample comprises 168 age-matched, sex-matched normal subjects from two studies: the Vanderbilt Memory and Aging Project (VMAP) and the Biomarkers of Cognitive Decline Among Normal Individuals (BIOCARD). First, we plotted the graph measures and used coefficient of variation (CoV) and the Mann-Whitney U test to evaluate different methods' effectiveness in removing site effects on the matrices and the derived graph measures. ComBat effectively eliminated site effects for global efficiency and modularity and outperformed the other two methods. However, all methods exhibited poor performance when harmonizing average betweenness centrality. Second, we tested whether our harmonization methods preserved correlations between age and graph measures. All methods except for CycleGAN in one direction improved correlations between age and global efficiency and between age and modularity from insignificant to significant with p-values less than 0.05.

Improved sensitivity and precision in multicentre diffusion MRI network analysis using thresholding and harmonization

PURPOSE

To investigate if network thresholding and raw data harmonization improve consistency of diffusion MRI (dMRI)-based brain networks while also increasing precision and sensitivity to detect disease effects in multicentre datasets.

METHODS

Brain networks were reconstructed from dMRI of five samples with cerebral small vessel disease (SVD; 629 patients, 166 controls), as a clinically relevant exemplar condition for studies on network integrity. We evaluated consistency of network architecture in age-matched controls, by calculating cross-site differences in connection probability and fractional anisotropy (FA). Subsequently we evaluated precision and sensitivity to disease effects by identifying connections with low FA in sporadic SVD patients relative to controls, using more severely affected patients with a pure form of genetically defined SVD as reference.

RESULTS

In controls, thresholding and harmonization improved consistency of network architecture, minimizing cross-site differences in connection probability and FA. In patients relative to controls, thresholding improved precision to detect disrupted connections by removing false positive connections (precision, before: 0.09-0.19; after: 0.38-0.70). Before harmonization, sensitivity was low within individual sites, with few connections surviving multiple testing correction (k = 0-25 connections). Harmonization and pooling improved sensitivity (k = 38), while also achieving higher precision when combined with thresholding (0.97).

CONCLUSION

We demonstrated that network consistency, precision and sensitivity to detect disease effects in SVD are improved by thresholding and harmonization. We recommend introducing these techniques to leverage large existing multicentre datasets to better understand the impact of disease on brain networks.

Multimodal tract-based MRI metrics outperform whole brain markers in determining cognitive impact of small vessel disease-related brain injury

In cerebral small vessel disease (cSVD), whole brain MRI markers of cSVD-related brain injury explain limited variance to support individualized prediction. Here, we investigate whether considering abnormalities in brain tracts by integrating multimodal metrics from diffusion MRI (dMRI) and structural MRI (sMRI), can better capture cognitive performance in cSVD patients than established approaches based on whole brain markers. We selected 102 patients (73.7 ± 10.2 years old, 59 males) with MRI-visible SVD lesions and both sMRI and dMRI. Conventional linear models using demographics and established whole brain markers were used as benchmark of predicting individual cognitive scores. Multi-modal metrics of 73 major brain tracts were derived from dMRI and sMRI, and used together with established markers as input of a feed-forward artificial neural network (ANN) to predict individual cognitive scores. A feature selection strategy was implemented to reduce the risk of overfitting. Prediction was performed with leave-one-out cross-validation and evaluated with the R² of the correlation between measured and predicted cognitive scores. Linear models predicted memory and processing speed with R² = 0.26 and R² = 0.38, respectively. With ANN, feature selection resulted in 13 tract-specific metrics and 5 whole brain markers for predicting processing speed, and 28 tract-specific metrics and 4 whole brain markers for predicting memory. Leave-one-out ANN prediction with the selected features achieved R² = 0.49 and R² = 0.40 for processing speed and memory, respectively. Our results show proof-of-concept that combining tract-specific multimodal MRI metrics can improve the prediction of cognitive performance in cSVD by leveraging tract-specific multi-modal metrics.

Cross-site harmonization of multi-shell diffusion MRI measures based on rotational invariant spherical harmonics (RISH)

Quantification methods based on the acquisition of diffusion magnetic resonance imaging (dMRI) with multiple diffusion weightings (e.g., multi-shell) are becoming increasingly applied to study the in-vivo brain. Compared to single-shell data for diffusion tensor imaging (DTI), multi-shell data allows to apply more complex models such as diffusion kurtosis imaging (DKI), which attempts to capture both diffusion hindrance and restriction effects, or biophysical models such as NODDI, which attempt to increase specificity by separating biophysical components. Because of the strong dependence of the dMRI signal on the measurement hardware, DKI and NODDI metrics show scanner and site differences, much like other dMRI metrics. These effects limit the implementation of multi-shell approaches in multicenter studies, which are needed to collect large sample sizes for robust analyses. Recently, a post-processing technique based on rotation invariant spherical harmonics (RISH) features was introduced to mitigate cross-scanner differences in DTI metrics. Unlike statistical harmonization methods, which require repeated application to every dMRI metric of choice, RISH harmonization is applied once on the raw data, and can be followed by any analysis. RISH features harmonization has been tested on DTI features but not its generalizability to harmonize multi-shell dMRI. In this work, we investigated whether performing the RISH features harmonization of multi-shell dMRI data removes cross-site differences in DKI and NODDI metrics while retaining longitudinal effects. To this end, 46 subjects underwent a longitudinal (up to 3 time points) two-shell dMRI protocol at 3 imaging sites. DKI and NODDI metrics were derived before and after harmonization and compared both at the whole brain level and at the voxel level. Then, the harmonization effects on cross-sectional and on longitudinal group differences were evaluated. RISH features averaged for each of the 3 sites exhibited prominent between-site differences in the frontal and posterior part of the brain. Statistically significant differences in fractional anisotropy, mean diffusivity and mean kurtosis were observed both at the whole brain and voxel level between all the acquisition sites before harmonization, but not after. The RISH method also proved effective to harmonize NODDI metrics, particularly in white matter. The RISH based harmonization maintained the magnitude and variance of longitudinal changes as compared to the non-harmonized data of all considered metrics. In conclusion, the application of RISH feature based harmonization to multi-shell dMRI data can be used to remove cross-site differences in DKI metrics and NODDI analyses, while retaining inherent relations between longitudinal acquisitions.

Determining the OPTIMAL DTI analysis method for application in cerebral small vessel disease

•

We were not able to identify a single optimal diffusion-weighted imaging analysis strategy across all 6 cohorts.

•

Diffusion tensor imaging measures at baseline predicted dementia conversion in cerebral small vessel disease and mild cognitive impairment.

•

Diffusion tensor imaging measures at baseline may be sensitive to differentiate between later vascular dementia vs Alzheimer’s disease dementia.

•

Diffusion tensor imaging measures significantly changed over time in cohorts with cerebral small vessel disease and cohorts with mild cognitive impairment. Change in diffusion tensor imaging measures were only consistently associated with dementia conversion in severe SVD.

•

The diffusion tensor imaging measures PSMD and DSEG required the lowest minimum sample sizes for a hypothetical clinical trial in patients with sporadic cerebral small vessel disease and mild cognitive impairment.

Background

DTI is sensitive to white matter (WM) microstructural damage and has been suggested as a surrogate marker for phase 2 clinical trials in cerebral small vessel disease (SVD). The study’s objective is to establish the best way to analyse the diffusion-weighted imaging data in SVD for this purpose. The ideal method would be sensitive to change and predict dementia conversion, but also straightforward to implement and ideally automated. As part of the OPTIMAL collaboration, we evaluated five different DTI analysis strategies across six different cohorts with differing SVD severity.

Methods

Those 5 strategies were: (1) conventional mean diffusivity WM histogram measure (MD median), (2) a principal component-derived measure based on conventional WM histogram measures (PC1), (3) peak width skeletonized mean diffusivity (PSMD), (4) diffusion tensor image segmentation θ (DSEG θ) and (5) a WM measure of global network efficiency (Geff). The association between each measure and cognitive function was tested using a linear regression model adjusted by clinical markers. Changes in the imaging measures over time were determined. In three cohort studies, repeated imaging data together with data on incident dementia were available. The association between the baseline measure, change measure and incident dementia conversion was examined using Cox proportional-hazard regression or logistic regression models. Sample size estimates for a hypothetical clinical trial were furthermore computed for each DTI analysis strategy.

Results

There was a consistent cross-sectional association between the imaging measures and impaired cognitive function across all cohorts. All baseline measures predicted dementia conversion in severe SVD. In mild SVD, PC1, PSMD and Geff predicted dementia conversion. In MCI, all markers except Geff predicted dementia conversion. Baseline DTI was significantly different in patients converting to vascular dementia than to Alzheimer’ s disease. Significant change in all measures was associated with dementia conversion in severe but not in mild SVD. The automatic and semi-automatic measures PSMD and DSEG θ required the lowest minimum sample sizes for a hypothetical clinical trial in single-centre sporadic SVD cohorts.

Conclusion

DTI parameters obtained from all analysis methods predicted dementia, and there was no clear winner amongst the different analysis strategies. The fully automated analysis provided by PSMD offers advantages particularly for large datasets.