Hyperbolic geometry offers a natural space for modeling hierarchical and multi‑scale structures. HypTI extends this idea to the tabular‑MRI fusion scenario. Clinical tabular variables are projected into Lorentz space via learnable Lorentz prototypes, which encode the hierarchy among assessments. MRI scans are processed with an ROI‑guided multi‑scale representation learner that embeds anatomical patterns across brain regions and spatial scales into the same Lorentz space. The two embeddings are merged by an adaptive fusion module that produces subject‑specific scale and shift coefficients, dynamically modulating and integrating the multimodal vectors to obtain a patient‑level representation for dementia detection. Experiments on the OASIS‑3 dataset show that HypTI outperforms several representative baselines in metrics such as accuracy and recall, confirming the benefit of hyperbolic space for multimodal medical data fusion.
Review