From Regions to Manifolds: linking Radiomic Signatures with PCA and UMAP Brain Representations in Glioma

Abstract

Surgical resection is the standard treatment for glioma patients, yet postoperative outcomes remain highly variable. In this study, we investigated structural brain reorganisation following surgery using a multiscale imaging approach. We extracted five key radiomic features from T1-weighted MRI scans within 116 anatomical regions and the tumor area and applied Support Vector Machine classification to identify regions predictive of shorter- versus longer-term survival. To characterise global structural variability, both Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP) were applied across patients to derive low-dimensional spatial representations. Region-based analysis highlighted peri-tumoral cortices, as well as subcortical and cerebellar structures, as most discriminative. At the global level, the application of Euclidean distance computed in the embedding spaces revealed more pronounced longitudinal changes in shorter-term survivors, whereas longer-term survivors exhibited comparatively more stable patterns over time. Overlaying radiomic regions on Euclidean distance maps from the PCA components indicated partial spatial correspondence, suggesting complementary information from local and global features. These results demonstrate the utility of combining regional radiomics with global manifold learning to characterise postoperative brain changes and their association with survival. This multiscale framework may inform future prognostic models and guide surgical and therapeutic strategies in glioma patients.