Cross-scanner harmonization methods for structural MRI may need further work: A comparison study.

No prior work has fully resolved the persistent challenge of scanner-induced bias in multi-site neuroimaging studies. Precise brain morphological measurements are essential for clinical dementia research. However, variations in hardware and protocols often introduce significant noise into structural data. This uncertainty drove the development of post-acquisition harmonization techniques. Prior research has shown that raw data often lack consistency across different manufacturers. That gap motivated the current evaluation of various correction algorithms. It was already known that scanner differences complicate longitudinal tracking of cortical thinning. This study addresses the limitations of existing approaches in mitigating these technical discrepancies.

The aim of this study is to evaluate the effectiveness of various post-acquisition harmonization methods for structural brain imaging. Researchers sought to determine if deep learning and statistical tools could mitigate noise introduced by scanner variations. The clinical utility of biomarkers for aging and dementia depends heavily on precise morphological measurements. However, protocol differences between GE and Siemens hardware often compromise data integrity. This study addresses the specific problem of scanner-induced bias in both cross-sectional and longitudinal datasets. The authors motivated this investigation by comparing multiple correction strategies, including neural style transfer and ComBat. No prior work had systematically compared these diverse approaches across both native and preprocessed scan types. This analysis provides a necessary assessment of current capabilities in achieving consistent multi-site neuroimaging data.

The review approach involved a comparative assessment of deep learning, histogram matching, and statistical techniques. Researchers utilized a crossover dataset of 113 participants scanned on both GE and Siemens platforms. A longitudinal cohort of 454 individuals provided data for evaluating scanner changeover effects. The team performed harmonization on both raw native and preprocessed scans. They applied FreeSurfer version 7.1.1 to quantify cortical thickness across various brain regions. Statistical validation included Kolmogorov-Smirnov tests to examine data distributions. The investigators calculated intra-class correlation to determine agreement levels within the crossover group. Finally, they computed annualized percent change to assess the longitudinal performance of each algorithm.

Key findings from the literature indicate that harmonization performance is highly region-dependent across all tested methods. Initial raw native scans showed the least agreement at the frontal pole with an intra-class correlation of 0.72. Preprocessed scans exhibited even lower agreement at the frontal pole, reaching only 0.54. Neural style transfer, CycleGAN, and histogram matching improved intra-class correlation values to above 0.81 at the caudal anterior cingulate. These same methods raised frontal pole agreement to above 0.67. Longitudinal raw native scans displayed significant over- and under-estimations of thickness due to scanner switching. ComBat successfully matched thickness distributions but failed to remove the effects of scanner changeover. CycleGAN and neural style transfer provided slightly better results than other methods for longitudinal transitions.

The authors suggest that current harmonization techniques demonstrate comparable performance across different brain regions. No single method consistently outperformed others in normalizing data across all regions of interest. The researchers propose that existing tools remain insufficient for fully resolving longitudinal scanner-related discrepancies. This synthesis indicates that region-specific performance is a primary characteristic of current harmonization strategies. The study implies that future development must address the limitations observed in both deep learning and statistical approaches. These findings provide a framework for evaluating future improvements in scan consistency. The evidence highlights the necessity for more robust algorithms to handle longitudinal data variations. Ultimately, the authors conclude that further work is required to achieve reliable cross-scanner data integration.