Pinto Maíra Siqueira, Anania Vincenzo, Paolella Roberto, Smekens Céline, Billiet Thibo, Janssens Thomas, den Dekker Arnold J, Sijbers Jan, Guns Pieter-Jan, Van Dyck Pieter
Department of Radiology, University Hospital Antwerp (UZA), Antwerp, Belgium.
imec-Vision Lab, University of Antwerp, Antwerp, Belgium.
Front Neurosci. 2025 Jun 3;19:1591169. doi: 10.3389/fnins.2025.1591169. eCollection 2025.
The structural integrity of brain white matter is commonly assessed using quantitative diffusion metric maps derived from diffusion MRI (dMRI) data. However, in multi-site, multi-scanner studies, variability across and within scanners presents challenges in ensuring consistent and comparable diffusion evaluations. This study assesses the effectiveness of ComBat-based harmonization algorithms in reducing intra- and inter-scanner variability in diffusion metrics such as FA, MD, AD, RD, MK, AK, and RK. Utilizing the B-Q MINDED dataset, which includes anatomical and dMRI data from 38 healthy adults scanned twice on two 3T MRI scanners (Siemens Healthineers PrismaFit and Siemens Healthineers Skyra) on the same day, we evaluated the NeuroCombat and LongCombat algorithms for harmonizing diffusion metrics. These harmonization methods effectively minimized both intra- and inter-scanner variability, highlighting their potential to improve consistency in multi-scanner diffusion analysis. Our findings suggest that NeuroCombat and LongCombat are recommended for harmonizing dMRI metric maps in clinical studies. Additionally, both algorithms applied in either ROI-based or voxel-wise configurations, significantly reduced variability, achieving levels comparable to scan-rescan variability intra-scanner. Nonetheless, the choice of harmonization algorithm and implementation should be tailored to the research question at hand. Moreover, the significant intra- and inter-subject variability on non-harmonized diffusion data demonstrated in this study reinforces the importance of harmonization strategies that address any sources of variability. By minimizing scanner-specific biases, the NeuroCombat and LongCombat harmonization algorithms enhance the reliability of diffusion biomarkers, enabling large-scale studies and more informed clinical decision-making in brain-related conditions.
脑白质的结构完整性通常使用从扩散磁共振成像（dMRI）数据得出的定量扩散指标图进行评估。然而，在多站点、多扫描仪研究中，扫描仪之间和内部的变异性给确保一致且可比的扩散评估带来了挑战。本研究评估了基于ComBat的归一化算法在减少诸如分数各向异性（FA）、平均扩散率（MD）、轴向扩散率（AD）、径向扩散率（RD）、平均峰度（MK）、轴向峰度（AK）和径向峰度（RK）等扩散指标的扫描仪内部和扫描仪之间变异性方面的有效性。利用B-Q MINDED数据集，该数据集包括38名健康成年人在同一天在两台3T磁共振成像扫描仪（西门子医疗PrismaFit和西门子医疗Skyra）上进行两次扫描的解剖学和dMRI数据，我们评估了NeuroCombat和LongCombat算法对扩散指标进行归一化。这些归一化方法有效地最小化了扫描仪内部和扫描仪之间的变异性，突出了它们在改善多扫描仪扩散分析一致性方面的潜力。我们的研究结果表明，在临床研究中，推荐使用NeuroCombat和LongCombat对dMRI指标图进行归一化。此外，两种算法无论是应用于基于感兴趣区域（ROI）还是体素级配置，都显著降低了变异性，达到了与扫描仪内部重扫变异性相当的水平。尽管如此，归一化算法的选择和实施应根据手头的研究问题进行调整。此外，本研究中未归一化的扩散数据显示出显著的个体间和个体内变异性，这强化了应对任何变异性来源的归一化策略的重要性。通过最小化特定于扫描仪的偏差，NeuroCombat和LongCombat归一化算法提高了扩散生物标志物的可靠性，使大规模研究以及在脑部相关疾病中做出更明智的临床决策成为可能。
