Empirical assessment of the assumptions of ComBat with diffusion tensor imaging.

Purpose: Diffusion tensor imaging (DTI) is a magnetic resonance imaging technique that provides unique information about white matter microstructure in the brain but is susceptible to confounding effects introduced by scanner or acquisition differences. ComBat is a leading approach for addressing these site biases. However, despite its frequent use for harmonization, ComBat's robustness toward site dissimilarities and overall cohort size have not yet been evaluated in terms of DTI. Approach: As a baseline, we match N=358 participants from two sites to create a "silver standard" that simulates a cohort for multi-site harmonization. Across sites, we harmonize mean fractional anisotropy and mean diffusivity, calculated using participant DTI data, for the regions of interest defined by the JHU EVE-Type III atlas. We bootstrap 10 iterations at 19 levels of total sample size, 10 levels of sample size imbalance between sites, and 6 levels of mean age difference between sites to quantify (i) ßAGE, the linear regression coefficient of the relationship between FA and age; (ii) Î³/f*, the ComBat-estimated site-shift; and (iii) Î´/f*, the ComBat-estimated site-scaling. We characterize the reliability of ComBat by evaluating the root mean squared error in these three metrics and examine if there is a correlation between the reliability of ComBat and a violation of assumptions. Results: ComBat remains well behaved for ßAGE when N>162 and when the mean age difference is less than 4 years. The assumptions of the ComBat model regarding the normality of residual distributions are not violated as the model becomes unstable. Conclusion: Prior to harmonization of DTI data with ComBat, the input cohort should be examined for size and covariate distributions of each site. Direct assessment of residual distributions is less informative on stability than bootstrap analysis. We caution use ComBat of in situations that do not conform to the above thresholds.

DeepN4: Learning N4ITK Bias Field Correction for T1-weighted Images.

T1-weighted (T1w) MRI has low frequency intensity artifacts due to magnetic field inhomogeneities. Removal of these biases in T1w MRI images is a critical preprocessing step to ensure spatially consistent image interpretation. N4ITK bias field correction, the current state-of-the-art, is implemented in such a way that makes it difficult to port between different pipelines and workflows, thus making it hard to reimplement and reproduce results across local, cloud, and edge platforms. Moreover, N4ITK is opaque to optimization before and after its application, meaning that methodological development must work around the inhomogeneity correction step. Given the importance of bias fields correction in structural preprocessing and flexible implementation, we pursue a deep learning approximation / reinterpretation of the N4ITK bias fields correction to create a method which is portable, flexible, and fully differentiable. In this paper, we trained a deep learning network "DeepN4" on eight independent cohorts from 72 different scanners and age ranges with N4ITK-corrected T1w MRI and bias field for supervision in log space. We found that we can closely approximate N4ITK bias fields correction with naïve networks. We evaluate the peak signal to noise ratio (PSNR) in test dataset against the N4ITK corrected images. The median PSNR of corrected images between N4ITK and DeepN4 was 47.96 dB. In addition, we assess the DeepN4 model on eight additional external datasets and show the generalizability of the approach. This study establishes that incompatible N4ITK preprocessing steps can be closely approximated by naïve deep neural networks, facilitating more flexibility. All code and models are released at https://github.com/MASILab/DeepN4 .

MidRISH: Unbiased harmonization of rotationally invariant harmonics of the diffusion signal.

Data harmonization is necessary for removing confounding effects in multi-site diffusion image analysis. One such harmonization method, LinearRISH, scales rotationally invariant spherical harmonic (RISH) features from one site ("target") to the second ("reference") to reduce confounding scanner effects. However, reference and target site designations are not arbitrary and resultant diffusion metrics (fractional anisotropy, mean diffusivity) are biased by this choice. In this work we propose MidRISH: rather than scaling reference RISH features to target RISH features, we project both sites to a mid-space. We validate MidRISH with the following experiments: harmonizing scanner differences from 37 matched patients free of cognitive impairment, and harmonizing acquisition and study differences on 117 matched patients free of cognitive impairment. We find that MidRISH reduces bias of reference selection while preserving harmonization efficacy of LinearRISH. Users should be cautious when performing LinearRISH harmonization. To select a reference site is to choose diffusion metric effect-size. Our proposed method eliminates the bias-inducing site selection step.

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.

Characterizing patterns of DTI variance in aging brains.

Background: As large analyses merge data across sites, a deeper understanding of variance in statistical assessment across the sources of data becomes critical for valid analyses. Diffusion tensor imaging (DTI) exhibits spatially varying and correlated noise, so care must be taken with distributional assumptions. Purpose: We characterize the role of physiology, subject compliance, and the interaction of subject with the scanner in the understanding of DTI variability, as modeled in spatial variance of derived metrics in homogeneous regions. Methods: We analyze DTI data from 1035 subjects in the Baltimore Longitudinal Study of Aging (BLSA), with ages ranging from 22.4 to 103 years old. For each subject, up to 12 longitudinal sessions were conducted. We assess variance of DTI scalars within regions of interest (ROIs) defined by four segmentation methods and investigate the relationships between the variance and covariates, including baseline age, time from the baseline (referred to as "interval"), motion, sex, and whether it is the first scan or the second scan in the session. Results: Covariate effects are heterogeneous and bilaterally symmetric across ROIs. Inter-session interval is positively related (p ≪ 0.001) to FA variance in the cuneus and occipital gyrus, but negatively (p ≪ 0.001) in the caudate nucleus. Males show significantly (p ≪ 0.001) higher FA variance in the right putamen, thalamus, body of the corpus callosum, and cingulate gyrus. In 62 out of 176 ROIs defined by the Eve type-1 atlas, an increase in motion is associated (p < 0.05) with a decrease in FA variance. Head motion increases during the rescan of DTI (Δµ = 0.045 millimeters per volume). Conclusions: The effects of each covariate on DTI variance, and their relationships across ROIs are complex. Ultimately, we encourage researchers to include estimates of variance when sharing data and consider models of heteroscedasticity in analysis. This work provides a foundation for study planning to account for regional variations in metric variance.

Predicting Age from White Matter Diffusivity with Residual Learning.

Imaging findings inconsistent with those expected at specific chronological age ranges may serve as early indicators of neurological disorders and increased mortality risk. Estimation of chronological age, and deviations from expected results, from structural magnetic resonance imaging (MRI) data has become an important proxy task for developing biomarkers that are sensitive to such deviations. Complementary to structural analysis, diffusion tensor imaging (DTI) has proven effective in identifying age-related microstructural changes within the brain white matter, thereby presenting itself as a promising additional modality for brain age prediction. Although early studies have sought to harness DTI's advantages for age estimation, there is no evidence that the success of this prediction is owed to the unique microstructural and diffusivity features that DTI provides, rather than the macrostructural features that are also available in DTI data. Therefore, we seek to develop white-matter-specific age estimation to capture deviations from normal white matter aging. Specifically, we deliberately disregard the macrostructural information when predicting age from DTI scalar images, using two distinct methods. The first method relies on extracting only microstructural features from regions of interest (ROIs). The second applies 3D residual neural networks (ResNets) to learn features directly from the images, which are non-linearly registered and warped to a template to minimize macrostructural variations. When tested on unseen data, the first method yields mean absolute error (MAE) of 6.11 ± 0.19 years for cognitively normal participants and MAE of 6.62 ± 0.30 years for cognitively impaired participants, while the second method achieves MAE of 4.69 ± 0.23 years for cognitively normal participants and MAE of 4.96 ± 0.28 years for cognitively impaired participants. We find that the ResNet model captures subtler, non-macrostructural features for brain age prediction.

DeepN4: Learning N4ITK Bias Field Correction for T1-weighted Images.

T1-weighted (T1w) MRI has low frequency intensity artifacts due to magnetic field inhomogeneities. Removal of these biases in T1w MRI images is a critical preprocessing step to ensure spatially consistent image interpretation. N4ITK bias field correction, the current state-of-the-art, is implemented in such a way that makes it difficult to port between different pipelines and workflows, thus making it hard to reimplement and reproduce results across local, cloud, and edge platforms. Moreover, N4ITK is opaque to optimization before and after its application, meaning that methodological development must work around the inhomogeneity correction step. Given the importance of bias fields correction in structural preprocessing and flexible implementation, we pursue a deep learning approximation / reinterpretation of the N4ITK bias fields correction to create a method which is portable, flexible, and fully differentiable. In this paper, we trained a deep learning network "DeepN4" on eight independent cohorts from 72 different scanners and age ranges with N4ITK-corrected T1w MRI and bias field for supervision in log space. We found that we can closely approximate N4ITK bias fields correction with naïve networks. We evaluate the peak signal to noise ratio (PSNR) in test dataset against the N4ITK corrected images. The median PSNR of corrected images between N4ITK and DeepN4 was 47.96 dB. In addition, we assess the DeepN4 model on eight additional external datasets and show the generalizability of the approach. This study establishes that incompatible N4ITK preprocessing steps can be closely approximated by naïve deep neural networks, facilitating more flexibility. All code and models are released at https://github.com/MASILab/DeepN4.

MidRISH: Unbiased harmonization of rotationally invariant harmonics of the diffusion signal.

Objective: Data harmonization is necessary for removing confounding effects in multi-site diffusion image analysis. One such harmonization method, LinearRISH, scales rotationally invariant spherical harmonic (RISH) features from one site ("target") to the second ("reference") to reduce confounding scanner effects. However, reference and target site designations are not arbitrary and resultant diffusion metrics (fractional anisotropy, mean diffusivity) are biased by this choice. In this work we propose MidRISH: rather than scaling reference RISH features to target RISH features, we project both sites to a mid-space. Methods: We validate MidRISH with the following experiments: harmonizing scanner differences from 37 matched patients free of cognitive impairment, and harmonizing acquisition and study differences on 117 matched patients free of cognitive impairment. Conclusion: MidRISH reduces bias of reference selection while preserving harmonization efficacy of LinearRISH. Significance: Users should be cautious when performing LinearRISH harmonization. To select a reference site is to choose diffusion metric effect-size. Our proposed method eliminates the bias-inducing site selection step.

Comparing voxel- and feature-wise harmonization of complex graph measures from multiple sites for structural brain network investigation of aging.

Complex graph theory measures of brain structural connectomes derived from diffusion weighted images (DWI) provide insight into the network structure of the brain. Further, as the number of available DWI datasets grows, so does the ability to investigate associations in these measures with major biological factors, like age. However, one key hurdle that remains is the presence of scanner effects that can arise from different DWI datasets and confound multisite analyses. Two common approaches to correct these effects are voxel-wise and feature-wise harmonization. However, it is still unclear how to best leverage them for graph-theory analysis of an aging population. Thus, there is a need to better characterize the impact of each harmonization method and their ability to preserve age related features. We investigate this by characterizing four complex graph theory measures (modularity, characteristic path length, global efficiency, and betweenness centrality) in 48 participants aged 55 to 86 from Baltimore Longitudinal Study of Aging (BLSA) and Vanderbilt Memory and Aging Project (VMAP) before and after voxel- and feature-wise harmonization with the Null Space Deep Network (NSDN) and ComBat, respectively. First, we characterize across dataset coefficients of variation (CoV) and find the combination of NSDN and ComBat causes the greatest reduction in CoV followed by ComBat alone then NSDN alone. Second, we reproduce published associations of modularity with age after correcting for other covariates with linear models. We find that harmonization with ComBat or ComBat and NSDN together improves the significance of existing age effects, reduces model residuals, and qualitatively reduces separation between datasets. These results reinforce the efficiency of statistical harmonization on the feature-level with ComBat and suggest that harmonization on the voxel-level is synergistic but may have reduced effect after running through the multiple layers of the connectomics pipeline. Thus, we conclude that feature-wise harmonization improves statistical results, but the addition of biologically informed voxel-based harmonization offers further improvement.

Implementation considerations for deep learning with diffusion MRI streamline tractography.

One area of medical imaging that has recently experienced innovative deep learning advances is diffusion MRI (dMRI) streamline tractography with recurrent neural networks (RNNs). Unlike traditional imaging studies which utilize voxel-based learning, these studies model dMRI features at points in continuous space off the voxel grid in order to propagate streamlines, or virtual estimates of axons. However, implementing such models is non-trivial, and an open-source implementation is not yet widely available. Here, we describe a series of considerations for implementing tractography with RNNs and demonstrate they allow one to approximate a deterministic streamline propagator with comparable performance to existing algorithms. We release this trained model and the associated implementations leveraging popular deep learning libraries. We hope the availability of these resources will lower the barrier of entry into this field, spurring further innovation.

Convolutional-recurrent neural networks approximate diffusion tractography from T1-weighted MRI and associated anatomical context.

Diffusion MRI (dMRI) streamline tractography is the gold-standard for in vivo estimation of white matter (WM) pathways in the brain. However, the high angular resolution dMRI acquisitions capable of fitting the microstructural models needed for tractography are often time-consuming and not routinely collected clinically, restricting the scope of tractography analyses. To address this limitation, we build on recent advances in deep learning which have demonstrated that streamline propagation can be learned from dMRI directly without traditional model fitting. Specifically, we propose learning the streamline propagator from T1w MRI to facilitate arbitrary tractography analyses when dMRI is unavailable. To do so, we present a novel convolutional-recurrent neural network (CoRNN) trained in a teacher-student framework that leverages T1w MRI, associated anatomical context, and streamline memory from data acquired for the Human Connectome Project. We characterize our approach under two common tractography paradigms, WM bundle analysis and structural connectomics, and find approximately a 5-15% difference between measures computed from streamlines generated with our approach and those generated using traditional dMRI tractography. When placed in the literature, these results suggest that the accuracy of WM measures computed from T1w MRI with our method is on the level of scan-rescan dMRI variability and raise an important question: is tractography truly a microstructural phenomenon, or has dMRI merely facilitated its discovery and implementation?

Characterizing Streamline Count Invariant Graph Measures of Structural Connectomes.

BACKGROUND: While graph measures are used increasingly to characterize human connectomes, uncertainty remains in how to use these metrics in a quantitative and reproducible manner. Specifically, there is a lack of community consensus regarding the number of streamlines needed to generate connectomes. PURPOSE: The purpose was to define the relationship between streamline count and graph-measure value, reproducibility, and repeatability. STUDY TYPE: Retrospective analysis of previously prospective study. POPULATION: Ten healthy subjects, 70% female, aged 25.3 ± 5.9 years. FIELD STRENGTH/SEQUENCE: A 3-T, T1-weighted sequences and diffusion-weighted imaging (DWI) with two gradient strengths (b-values = 1200 and 3000 sec/mm2 , echo time [TE] = 68 msec, repetition time [TR] = 5.4 seconds, 120 slices, field of view = 188 mm2 ). ASSESSMENT: A total of 13 graph-theory measures were derived for each subject by generating probabilistic whole-brain tractography from DWI and mapping the structural connectivity to connectomes. The streamline count invariance from changes in mean, repeatability, and reproducibility were derived. STATISTICAL TESTS: Paired t-test with P value <0.05 was used to compare graph-measure means with a reference, intraclass correlation coefficient (ICC) to measure repeatability, and concordance correlation coefficient (CCC) to measure reproducibility. RESULTS: Modularity and global efficiency converged to their reference mean with ICC > 0.90 and CCC > 0.99. Edge count, small-worldness, randomness, and average betweenness centrality converged to the reference mean, with ICC > 0.90 and CCC > 0.95. Assortativity and average participation coefficient converged with ICC > 0.75 and CCC > 0.90. Density, average node strength, average node degree, characteristic path length, average local efficiency, and average clustering coefficient did not converge, though had ICC > 0.90 and CCC > 0.99. For these measures, alternate definitions that converge a reference mean are provided. DATA CONCLUSION: Modularity and global efficiency are streamline count invariant for greater than 6 million and 100,000 streamlines, respectively. Density, average node strength, average node degree, characteristic path length, average local efficiency, and average clustering coefficient were strongly dependent on streamline count. EVIDENCE LEVEL: 1. TECHNICAL EFFICACY: Stage 1.