Separating the contributions of brain compartments with spatial statistics

Le résumé fourni par la source

Magnetic Resonance Spectroscopic Imaging (MRSI) enables non-invasive mapping of brain metabolism, offering valuable insights into tissue biochemistry in both healthy and pathological conditions. However, its clinical interpretation remains limited by partial volume effects, as the relatively large voxel size leads to the mixing of multiple anatomical compartments within a single measurement. This overlap results in heterogeneous metabolic signals that are difficult to attribute to specific tissue types. To overcome this limitation, we propose a statistical framework designed to separate anatomical contributions within MRSI voxels. The method builds upon the Geographically Weighted Regression (GWR) framework, adapted to model metabolite concentrations as a function of predefined anatomical segmentations. This approach enables spatially localized estimation of tissue-specific metabolic profiles. Furthermore, to account for spatial variability in spectral quality across the MRSI grid, we introduce a progressive weighting scheme based on signal-to-noise ratio (SNR) and full width at half maximum (FWHM). This strategy replaces conventional binary thresholding, allowing for a more nuanced integration of data quality into the model. The method was evaluated on a cohort of healthy volunteers to assess its general performance. The model demonstrated high goodness-of-fit across most metabolites (R² ranging from 0.81 to 0.87), with lower performance observed for lactate (R² = 0.37), likely due to its low concentration and higher variability in healthy tissue. In addition, the proposed weighting scheme significantly improved robustness to spectral degradation, reducing the median relative error from 3.06% to 1.11% and the maximum error from 96.6% to 36.6%. The framework was further applied to a glioblastoma case, where strong performance was observed for all metabolites (R² > 0.90). In this clinical application, the model enabled a clear separation of metabolic profiles between tumor tissue, necrotic regions, and healthy tissue. It also revealed spatially heterogeneous metabolic patterns within pathological regions, highlighting its ability to capture intra-tissue variability. These findings support the relevance of the proposed approach for improving the interpretability of MRSI data in complex pathological contexts. Overall, this framework provides a physiologically meaningful representation of brain metabolism by explicitly accounting for partial volume effects and anatomical heterogeneity. The progressive weighting strategy enhances robustness to variable data quality, while the model's interpolation capability enables retrospective alignment of voxel-wise measurements with regions of interest. By providing both local and global quality metrics, the method ensures transparency and supports its potential integration into routine clinical workflows.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.