Separation of Anatomical Contributions in MRSI Using a Non-Negative Geographically Weighted Regression Model

Magnetic resonance spectroscopic imaging (MRSI) is a non-invasive neuroimaging tool that maps brain metabolites, providing insights into brain physiology and pathology. However, clinical adoption remains limited due to low spatial resolution and partial volume effects, where acquired signals detect metabolism from different tissue types, preventing clear identification of tissue-specific metabolic patterns. We developed a statistical framework based on non-negative geographically weighted regression (NNGWR) designed to overcome these limitations. The method integrates tissue partial volumes with MRSI data using spatially adaptive regression to interpolate metabolite levels across the imaging volume while separating distinct tissue contributions. Importantly, our model accounts for spatial heterogeneity within tissues rather than assuming metabolic uniformity, providing more physiologically realistic estimates. Simulations using synthetic ground truth data show that our method consistently achieves a lower median absolute percentage error than cubic spline and zero-filling interpolations, especially under high noise or low-resolution conditions. A total of 120 MRSI datasets were acquired at 3 T from a cohort of 36 healthy volunteers to assess predictive performance. To handle real-world data, we implemented a robust reweighting scheme that addresses variable spectral quality by progressively downweighting degraded data rather than discarding it, balancing data retention with model stability. Application to a glioblastoma clinical case demonstrated successful separation of metabolite ratios across healthy, tumour and necrotic tissues. The model identified intratumoural metabolic heterogeneity and provided local confidence for transparent clinical interpretation. By decomposing metabolic contributions by tissue type and accounting for natural spatial variations, the approach enables improved brain tumour characterisation. This model is applicable to other neurological conditions where tissue-specific metabolic alterations serve as disease biomarkers, helping both clinical diagnostics and research into brain pathophysiology.