SPD Matrix Learning for Neuroimaging Analysis: Perspectives, Methods, and Challenges

Neuroimaging provides essential tools for characterizing brain activity and inter-regional connectivity through modalities that capture complementary aspects of brain organization. At the same time, extracting meaningful neural signatures remains challenging because each modality introduces its own sources of variability, including measurement noise, spatial and temporal distortions, heterogeneous acquisition protocols, and limited sample sizes. A unifying perspective arises when these data are represented by symmetric positive definite (SPD) matrices, such as covariance, connectivity, diffusion, or deformation tensors, which encode second-order statistical relationships or local anatomical structure depending on the modality. Equipping the SPD space with Riemannian geometry yields a non-Euclidean framework for principled statistical analysis and machine learning on these representations. This review organizes learning methodologies on SPD manifolds under a unified framework termed SPD matrix learning. From this perspective, SPD modeling provides a common language across neuroimaging modalities and links modern learning pipelines to decades of geometric statistics. We show that (i) learning on the SPD manifold is mathematically well grounded and preserves key structural constraints such as symmetry and positive definiteness; (ii) SPD matrix learning encompasses a broad class of geometric statistical tools that have long been used in neuroimaging; and (iii) it is increasingly compatible with modern AI paradigms, thereby enabling new classes of neuroimaging problems involving manifold-valued data. Taken together, SPD matrix learning offers a unified and forward-looking framework for neuroimaging analysis.