The impact of B1+ inhomogeneity on image quality metrics and morphometric statistical inferences at 7 T MRI

Technical Summary: The Impact of B1+ Inhomogeneity on Image Quality and Morphometric Inferences at 7 T MRI

Problem Statement
Structural neuroimaging at ultra-high field (7 Tesla) offers superior spatial resolution and signal-to-noise ratio compared to lower field strengths, enabling precise characterization of cortical and subcortical anatomy. However, 7 T imaging is inherently susceptible to transmit field (B1+) inhomogeneity, which causes spatially varying flip angles and non-uniform image intensity. While the Magnetization Prepared 2 Rapid Acquisition Gradient Echoes (MP2RAGE) sequence mitigates some biases (such as receive field and spin density variations), residual B1+ inhomogeneity persists. This residual bias is particularly problematic in clinical populations (e.g., aging or those with mitochondrial mutations) where anatomical and physiological factors may exacerbate image quality issues. The extent to which residual B1+ inhomogeneity affects automated image quality control (QC) metrics and, crucially, the statistical validity of morphometric case-control studies remains insufficiently understood. Furthermore, it is unclear whether modern deep learning-based segmentation frameworks are robust enough to render explicit B1+ correction unnecessary.

Methodology
The study utilized submillimeter 7 T MP2RAGE brain acquisitions from two groups: 22 carriers of the m.3243A>G mitochondrial gene mutation (without severe acute phenotypes) and 15 age- and sex-matched healthy controls.

• Data Processing: B1+ maps were acquired using the SA2RAGE sequence. MP2RAGE "uniform" images were processed both before and after B1+ inhomogeneity correction using established MATLAB-based methods.
• Automated Quality Control (QC): Image quality was assessed using multiple automated pipelines:
  • MRIQC: A "Stanford RF" classifier and component metrics (e.g., intensity non-uniformity).
  • CAT (Computational Anatomy Toolbox): Structural Image Quality Ratings (SIQR) and component metrics.
  • Topological Proxies: FreeSurfer-derived Arithmetic Mean of Hemispheric Euler Numbers (AMHEN) and FastSurfer-derived Normative Outlier Numbers (NON).
• Morphometric Analysis: Case-control differences (carriers vs. controls) were analyzed using both registration-based (CAT) and deep learning-based (FastSurfer, SynthSeg, CerebNet) segmentation frameworks. Metrics included regional gray matter volumes and mean cortical thickness across multiple atlases (DK40, Destrieux, AAL3).
• Statistical Analysis: Group differences were evaluated using permutation-based inference (PALM) with age, sex, and intracranial volume as covariates.

Key Contributions

1. Evaluation of Automated QC at 7 T: The study systematically tested whether current automated QC tools (trained largely on 3 T data) can reliably detect B1+ inhomogeneity effects in 7 T MP2RAGE data.
2. Method-Dependent Impact: It demonstrates that the influence of B1+ correction is not uniform but highly dependent on the specific morphometric pipeline (registration-based vs. deep learning) and the metric used (volume vs. thickness).
3. Clinical Relevance: By including a clinical cohort with subtle, diffuse structural alterations, the study provides a stringent test of pipeline sensitivity that goes beyond healthy control validations.

Results

• Image Quality Metrics:
  • Binary Classifiers: Pretrained binary classifiers (MRIQC "Stanford RF" and "Oxford RUS") failed to consistently distinguish between corrected and uncorrected images, often misclassifying uncorrected data as high quality or failing to process specific uncorrected carrier scans.
  • Component Metrics: Specific metrics sensitive to bias fields showed significant improvement. The intensity non-uniformity metric (inu_med) decreased significantly, and the topological proxy (AMHEN) increased significantly after correction, indicating fewer reconstruction defects.
  • Deep Learning Robustness: The FastSurfer-derived Normative Outlier Number (NON) did not change significantly, suggesting that deep learning reconstruction pipelines are robust to residual intensity inhomogeneity, even when the underlying image quality is technically improved.
• Morphometric Statistical Inferences:
  • Sensitivity Changes: B1+ correction significantly altered the number and spatial distribution of statistically significant regions. In many pipelines (e.g., CAT and FastSurfer cortical thickness), correction increased statistical sensitivity, revealing significant group differences that were previously undetectable.
  • Redistribution of Findings: In volumetric analyses (e.g., CAT AAL3), correction led to a redistribution of significant findings, with some regions losing significance while others gained it.
  • Pipeline Variability: Deep learning-based pipelines (SynthSeg, CerebNet) showed greater stability in the set of significant regions compared to registration-based pipelines, yet they still exhibited changes in sensitivity and significance levels after correction.

Significance and Conclusions
The paper concludes that residual B1+ inhomogeneity at 7 T is a critical confound that influences not only image appearance but also automated QC outcomes and, most importantly, the statistical inferences drawn from morphometric studies.

• QC Limitations: Current automated QC approaches are insufficient for reliably capturing B1+ inhomogeneity effects at 7 T; reliance on binary quality labels may be misleading.
• Necessity of Correction: Explicit B1+ inhomogeneity correction is practically necessary and highly recommended for establishing reliable morphometric biomarkers at ultra-high field. Even though deep learning pipelines offer robustness, they do not fully compensate for uncorrected bias, and correction can fundamentally alter neurobiological conclusions.
• Validation Strategy: The study emphasizes that validation of neuroimaging pipelines in healthy populations is insufficient for clinical applications. Clinical cohorts with subtle pathologies are essential to reveal hidden sensitivities in preprocessing strategies.

The authors assert that to ensure reproducible and biologically meaningful results at 7 T, researchers must implement explicit B1+ correction and carefully select preprocessing strategies tailored to the specific analytical framework and population.