Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster D in Poster Session D: Wednesday, August 5, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

UMBRELLA: A Foundation Framework for High-Dimensional Neuroimaging via Multi-turn Comparative Learning

Heehwan Wang¹, Yujin Yun¹, Suin Cho², Jiook Cha¹; ¹Seoul National University, ²Boston University

Presenter: Heehwan Wang

Deep learning for neuroimaging is limited by data scarcity and high computational costs. We introduce UMBRELLA, a framework adapting pretrained Vision-Language Models (VLMs) via 3D patch embedding and a frozen backbone. To overcome data constraints, we propose Comparison-based Learning, leveraging the VLM's native interleaving mechanism to contrast query scans against normative references. Evaluations on the ABCD dataset show that this relational approach outperforms single-scan baselines and scales with the number of reference images. Furthermore, multimodal integration (sMRI+dMRI) preserves performance across individual modalities. Our results demonstrate that repurposing VLM interleaving for relational analysis provides a scalable, computationally efficient path toward neuroimaging foundation models.

Topic Area: Methods, Tools, Theory & Neural Coding