The utility of explainable AI for MRI analysis: Relating model predictions to neuroimaging features of the aging brain

S. Hofmann, Ole Goltermann, Nico Scherf, Klaus-Robert Müller, Markus Loeffler, Arno Villringer, Michael Gaebler, A. Veronica Witte, Frauke BeyerImaging 2025 · DOI

TL;DR

This study bridges the gap between highly accurate deep learning models for predicting brain age from MRI and the "black box" problem of understanding why they make those predictions. By using explainable AI (XAI) to highlight important brain regions and then correlating these "relevance maps" with known aging biomarkers, the researchers found that these AI models effectively capture established patterns of brain aging, even highlighting the cerebellum's significant role more than expected.

What Problem Does This Paper Address?

Deep learning (DL) models have become remarkably good at predicting a person's "brain age" from MRI scans, often outperforming traditional methods. This brain age is a measure of how biologically old a brain appears, and deviations from chronological age can indicate accelerated or decelerated aging. However, DL models are notoriously opaque; we don't know which specific brain features they are using to make these age predictions. Explainable AI (XAI) techniques can generate "relevance maps" that highlight the voxels (3D pixels) in an MRI that most influenced the model's prediction. While these maps tell us where the model is looking, they don't tell us what biological features (like tissue volume, thickness, or integrity) are present in those highlighted areas. This paper aims to close that gap by systematically linking XAI-derived relevance maps from DL-based brain age prediction models to established neuroimaging markers of aging.

Methods

The researchers employed a multi-level ensemble of 3D convolutional neural networks (3D-CNNs) trained on T1-weighted and FLAIR MRI sequences from 1855 participants. To understand the models' decision-making process, they utilized Layer-wise Relevance Propagation (LRP), a popular XAI technique. LRP decomposes the model's prediction back through its layers to assign a "relevance score" to each input voxel, indicating its contribution to the predicted brain age.

The core of the study involved correlating these LRP-generated relevance maps with various neuroimaging markers of brain aging. These markers included:

  • Cortical thickness and gray matter volume (GMV): Measured using FreeSurfer.
  • Subcortical, brainstem, and cerebellar volumes: Also derived from FreeSurfer.
  • Ventricular volumes: Including lateral and third ventricles.
  • Perivascular Spaces (PVS): Detected using a deep learning-based toolbox (SHIVA-PVS).
  • Fractional Anisotropy (FA): A measure of white matter integrity derived from diffusion-weighted MRI (DWI).

Statistical analyses, primarily Pearson correlations (with Bonferroni correction for multiple comparisons), were used to quantify the relationships between the XAI relevance scores and these imaging features.

Datasets

The study utilized data from the LIFE-Adult study, a large, population-based cohort in Leipzig, Germany. This cohort provided extensive clinical and neuroimaging data. Specifically, MRI data from 1855 participants (aged 18-82 years) were used for the analysis. No new datasets were introduced; the focus was on analyzing existing data with novel XAI-to-biomarker correlation methods.

Key Insights and Ideas

  • XAI as a Bridge: The study successfully demonstrated that XAI methods can act as a crucial bridge, translating the "black box" outputs of deep learning models into biologically meaningful insights about brain aging.
  • Cerebellum's Underappreciated Role: The cerebellum, often considered secondary in aging studies compared to cortical regions, emerged as a significant contributor to brain age prediction, showing strong correlations between its gray matter volume and model relevance.
  • PVS Not a Key Driver (for these models): Contrary to expectations, perivascular spaces (PVS), a marker of small vessel disease, were not consistently identified as relevant by the DL models, suggesting they might be too subtle or confounded by other features for these specific models.
  • Ventricular Expansion is a Strong Signal: The expansion of ventricles, a well-known sign of brain aging, showed the strongest correlation with model relevance, highlighting its importance for age prediction.
  • White Matter Integrity Matters: Measures of white matter integrity (FA) were also significantly correlated with model relevance, particularly in frontal and occipital regions, aligning with known age-related white matter changes.

Key Results

  • Broad Correlation: All neuroimaging markers of brain aging, except for PVS, showed significant correlations with the XAI-based relevance maps.
  • Strongest Correlation: Ventricular volume had the strongest correlation with relevance (r = 0.69).
  • Feature-Specific Strong Correlations: Temporal-parietal cortical thickness and volume, cerebellar gray matter volume, and frontal-occipital white matter tracts showed particularly strong correlations with XAI-based relevance.
  • Cortical Aging: 42 out of 68 cortical regions showed a statistically significant negative correlation between relevance scores and cortical thickness (mean r = -0.20), and 44 out of 68 regions showed a negative correlation with GMV (mean r = -0.22).
  • Subcortical Aging: 19 out of 25 subcortical regions showed a significant negative correlation between relevance scores and GMV (mean r = -0.28), with the strongest being in the left cerebellar GMV (r = -0.52).
  • White Matter: In region-based analysis, 23 out of 48 WM regions showed a negative correlation between relevance and FA (mean r = -0.13), with the strongest in the left superior fronto-occipital fasciculus (r = -0.25).

Limitations and Future Work

  • PVS Confounding: The lack of consistent relevance for PVS might be due to their close proximity to WMH, which are more readily detected by FLAIR images, or because PVS are fluid-filled and have zero intensity in T1w images, making them less directly interpretable by the models.
  • Model Specificity: The findings are specific to the DL models and XAI methods used. Different model architectures or XAI techniques might yield different results.
  • Limited Comparison: The study did not directly compare DL+XAI with traditional neuroimaging analysis methods, focusing instead on validating the XAI insights.
  • Future Work:
    • Investigate the role of DTI data directly in DL model training for more comprehensive WM analysis.
  • Explore advanced XAI techniques (e.g., Concept Relevance Propagation) to capture interactions between brain regions.
  • Further explore the cognitive implications of high brain age predictions in specific regions, particularly the cerebellum.
  • Develop methods to identify unknown brain changes not yet captured by current biomarkers.

Summary updated

Abstract

Deep learning models highly accurately predict brain age from MRI but their explanatory capacity is limited. Explainable artificial intelligence (XAI) methods can identify relevant voxels contributing to model estimates, yet they do not reveal which biological features these voxels represent. In this study, we closed this gap by relating voxel-based contributions to brain-age estimates, extracted with XAI, to human-interpretable structural features of the aging brain. To this end, we associated participant-level XAI-based relevance maps extracted from two ensembles of 3D-convolutional neural networks (3D-CNNs) that were trained on T1-weighted and fluid-attenuated inversion recovery images of 1855 participants (age range 18-82 years), with regional cortical and subcortical gray matter volume and thickness, perivascular spaces (PVS), and water diffusion-based fractional anisotropy of major white matter tracts. We found that all neuroimaging markers of brain aging, except for PVS, were highly correlated with the XAI-based relevance maps. Overall, the strongest correlation was found between ventricular volume and relevance (<i>r</i>= 0.69), and by feature, temporal-parietal cortical thickness and volume, cerebellar gray matter volume, and frontal-occipital white matter tracts showed the strongest correlations with XAI-based relevance. Our ensembles of 3D-CNNs took into account a plethora of known aging processes in the brain to perform age prediction. Some age-associated features like PVS were not consistently considered by the models, and the cerebellum was more important than expected. Taken together, we highlight the ability of end-to-end deep learning models combined with XAI to reveal biologically relevant, multi-feature relationships in the brain.
BibTeX
@inproceedings{hofmann2025utility,
  author = {S. Hofmann and Ole Goltermann and Nico Scherf and Klaus-Robert Müller and Markus Loeffler and Arno Villringer and Michael Gaebler and A. Veronica Witte and Frauke Beyer},
  title = {{The utility of explainable AI for MRI analysis: Relating model predictions to neuroimaging features of the aging brain}},
  booktitle = {Imaging},
  year = {2025},
  doi = {10.1162/imag\_a\_00497},
}
Figure