(E-172) Evaluating transferability of ComBat harmonization of diffusion tensor magnetic resonance imaging data

Introduction: Multisite magnetic resonance imaging (MRI) studies enable achievement of greater study power over a shorter duration of time, and are becoming increasingly prevalent in neuroimaging research. However, pooling of data from multiple sites is limited by noise characteristics of individual scanners and their receive chain elements (e.g., coils, filters, algorithms). Data harmonization is desired to minimize the biases introduced by such limitations. One such algorithm, ComBat [1], has been demonstrated to be effective at reducing scanner-specific variability in neuroimaging measures derived from multisite MRI [2], [3]. Despite increased usage in neuroimaging research [4], the ComBat algorithm retains the shortcoming that one must re-compute the harmonizing transformation after adding any new subjects to the data pool. This is undesirable for larger, long-term multisite collaborations where multiple analyses may need to be conducted over years of data collection. Solutions have been briefly explored by Da-ano et al. [5], who determined that ComBat could be used to compute a transferable harmonizing transformation of multicenter radiomics features, which could be reapplied to harmonize new subjects from a known imaging center, albeit with slightly decreased performance. The key benefit of this transferable ComBat methodology (hereafter referred to as T-ComBat) is that the computed transformation may be used to incorporate new subject data into the harmonized space without affecting prior data. Here we evaluate the transferability of ComBat data harmonization to previously unseen subjects as applied to two diffusion tensor imaging (DTI) MRI metrics: fractional anisotropy (FA) and mean diffusivity (MD).


Materials and Methods: Imaging: 10 subjects (5 male; ages 23-30) underwent four MRI scan sessions. Two sessions (GE1, GE2) were conducted on a 3T GE Discovery MR750 with 32-channel brain array (Nova Medical). Two sessions (S1, S2) were conducted on a 3T Siemens Prisma with 64-channel brain array. For all subjects (except one, whose GE1 and S1 were conducted on separate dates), GE1 and S1 were conducted during a single visit with GE2 and S2 conducted during a second. In every session, a T1-weighted brain volume and 64-direction DW-MRI (EPI, 2mm isotropic, FOV = 24x24cm) were collected. After Eddy current and motion correction (FSL eddy_cuda9.1 [6]–[8]), DW-MRI images were used to compute voxel-wise FA and MD maps (FSL dtifit) which were registered (ANTs antsRegistrationSyNQuick.sh [9]) to a standard atlas [10].

Analysis: Subject-wise root mean square error (RMSE) comparing FA maps across session pairings (GE1-S1, GE2-S2) was computed and averaged over all subjects. For GE1 and S1, each subject’s voxel-wise FA was averaged within an eroded white matter (WM) mask (Figure 1a), producing FA_WM. The Wilcoxon signed rank test evaluated significance of differences between subject-wise GE1 and S1 FA_WM. Metrics were computed for three cases: (1) unharmonized FA maps, (2) after traditional ComBat, in which all subjects’ data were used to conduct harmonization, and (3) after T-ComBat, in which the test subject was excluded from data used to compute ComBat harmonization parameters which were saved and used to harmonize the test subject’s data (repeated for each subject). Identical analyses were conducted on MD data.


Results, Conclusions, and Discussions: Results: Full ComBat and T-ComBat both reduced differences in FA and MD maps across scanners. Averaged RMSE_FA and RMSE_MD for the unharmonized, full ComBat, and T-ComBat data are shown in Table 1. Statistical testing of GE1 versus S1 FA_WM resulted in a significant difference in the unharmonized data (p=0.002), but not after full ComBat or T-ComBat (p>0.05). Similarly, analysis of GE1 versus S1 MD_WM resulted in a significant difference only for the unharmonized data (p=0.002). Boxplots of subject-wise difference between GE1 and S1 FA_WM (ΔFA_WM,GE1-S1) are shown in Figure 1b for the original (unharmonized) data and the two harmonization methods. Similarly, ΔMD_WM,GE1-S1 is presented in Figure 1c.

Discussion: T-ComBat represents a cross-validation approach to assessing whether ComBat can be used to compute a harmonizing transformation based on a set of multi-scanner “training” data that is then reapplied to harmonize a new, previously-unseen “test” subject. Given that it successfully eliminated statistical significance in cross-scanner differences in FA_WM and MD_WM, T-ComBat may be a viable method for harmonization of new subjects that does not require adjustment of previously included data. We note, however, that the WM mask used was strongly eroded (to ensure that the regions included in the analysis represent WM for all subjects), limiting these results primarily to deeper areas of WM tissue. Further, the averaged RMSE_FA and RMSE_MD metrics (Table 1) demonstrate that while FA and MD map similarity across scanners was improved under the T-ComBat approach, it did not reach the performance level achieved by full ComBat. T-ComBat achieved roughly 74% of the performance of full ComBat for FA harmonization, and roughly 56% for MD harmonization. The current results are also limited by the small sample size of the dataset.

Conclusion: We demonstrated that T-ComBat successfully removes statistically significant differences between inter-scanner averaged FA and MD within WM, but does not achieve the harmonizing performance of ComBat applied to the full dataset. This highlights the potential of the transferable ComBat method as a means for harmonizing previously unseen subject data across MRI scanners, though further evaluation on a larger dataset is required to better understand the method’s limitations.


Acknowledgements (Optional) : This work was supported by a Core Facility grant from the Indiana Clinical and Translational Sciences Institute (PIs: Thomas Talavage & Joaquín Goñi). This work was also supported by the National Institutes of Health grant: S10 OD012336 – 3T MRI Scanner Dedicated to Life Sciences Research (PI: Ulrike Dydak).




References (Optional) : [1] W. E. Johnson, C. Li, and A. Rabinovic, “Adjusting batch effects in microarray expression data using empirical Bayes methods,” Biostat. Oxf. Engl., vol. 8, no. 1, pp. 118–127, Jan. 2007, doi: 10.1093/biostatistics/kxj037.

[2] J.-P. Fortin et al., “Harmonization of multi-site diffusion tensor imaging data,” NeuroImage, vol. 161, pp. 149–170, Nov. 2017, doi: 10.1016/j.neuroimage.2017.08.047.

[3] J.-P. Fortin et al., “Harmonization of cortical thickness measurements across scanners and sites,” NeuroImage, vol. 167, pp. 104–120, Feb. 2018, doi: 10.1016/j.neuroimage.2017.11.024.

[4] F. Orlhac et al., “A guide to ComBat harmonization of imaging biomarkers in multicenter studies,” J. Nucl. Med., Sep. 2021, doi: 10.2967/jnumed.121.262464.

[5] R. Da-ano et al., “A transfer learning approach to facilitate ComBat-based harmonization of multicentre radiomic features in new datasets,” PLOS ONE, vol. 16, no. 7, p. e0253653, Jul. 2021, doi: 10.1371/journal.pone.0253653.

[6] M. Jenkinson, C. F. Beckmann, T. E. J. Behrens, M. W. Woolrich, and S. M. Smith, “FSL,” NeuroImage, vol. 62, no. 2, pp. 782–790, Aug. 2012, doi: 10.1016/j.neuroimage.2011.09.015.

[7] S. M. Smith et al., “Advances in functional and structural MR image analysis and implementation as FSL,” NeuroImage, vol. 23, pp. S208–S219, Jan. 2004, doi: 10.1016/j.neuroimage.2004.07.051.

[8] M. W. Woolrich et al., “Bayesian analysis of neuroimaging data in FSL,” NeuroImage, vol. 45, no. 1, Supplement 1, pp. S173–S186, Mar. 2009, doi: 10.1016/j.neuroimage.2008.10.055.

[9] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee, “Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain,” Med. Image Anal., vol. 12, no. 1, pp. 26–41, Feb. 2008, doi: 10.1016/j.media.2007.06.004.

[10] V. Fonov, A. C. Evans, K. Botteron, C. R. Almli, R. C. McKinstry, and D. L. Collins, “Unbiased average age-appropriate atlases for pediatric studies,” NeuroImage, vol. 54, no. 1, pp. 313–327, Jan. 2011, doi: 10.1016/j.neuroimage.2010.07.033.