Reducing MRI Inter-Scanner Variability Using 3D Superpixel ComBat

Background: Inter-scanner variability hinders the direct comparability of multi-site/scanner MRI data for clinical research. The ComBat method is commonly used to reduce the variability based on an empirical Bayes framework^1,2, harmonizing the data at the feature level (e.g., region-of-interest measures). However, directly harmonizing the scans at the voxel-level using ComBat has been relatively less explored. In this study, we investigated the performance of the voxel-wise ComBat. Also, going beyond voxels, we proposed a new ComBat approach which operates on a small group of voxels called superpixels³. Method: Eighteen subjects (10 patients with Alzheimer's disease and 8 controls; age: 68.0 [9.3] years; 10 females) participated in this study. For each subject, T1-weighted images were acquired on each of four 3T scanners with different manufacturers or models (i.e., GE, Philips, Siemens-Prisma, Siemens-Trio). After the standard image preprocessing including two-step registration by using the Statistical Parametric Mapping (SPM12)⁴, the unharmonized scans (Raw data) were aligned in the standard template space. To reduce the computational load for ComBat at the voxel level (Voxel-ComBat), we used a three-dimensional superpixel algorithm³ to parcellate the images into hundreds of superpixels based on the study-specific template, and then the ComBat was applied at the superpixel level (Figure 1). Compared to Voxel-Combat operating on about half million voxels (computation time >>10,000 seconds), this superpixel ComBat (SP-ComBat) operates on only a few hundred superpixels, significantly improving the computation efficiency (computation time < 5 seconds) while maintaining the harmonization performance. The harmonized scans were used to estimate cortical thickness by employing surface-based morphometry⁵, and the coefficients of variation of thickness measures were calculated to evaluate the harmonization performance. Result: The harmonized data provided similar contrasts across scanners compared to the Raw images in visual inspection (Figure 2) and had comparable distributions of the tissue-specific signal intensity between scanners for both Voxel-ComBat and SP-ComBat (Figure 3). Also, these two methods significantly reduced the inter-scanner variation (both p-values < 0.001) in terms of cortical thickness measures (Figure 4). Conclusion: This study evaluated the feasibility and effectiveness of Voxel-ComBat and proposed a new approach, SP-ComBat, to optimize the efficiency of ComBat harmonization at the voxel level.