Schedule of Events | Search Abstracts | Invited Symposia | Symposia | Poster Sessions | Data Blitz

Automated network labeling with contrastive neuroimage–language models

Poster Session B - Sunday, March 8, 2026, 8:00 – 10:00 am PDT, Fairview/Kitsilano Ballroom

Ryan Hammonds¹ (rphammonds@ucsd.edu), Jerjes Aguirre-Chavez¹, Borngreat Omoma-Edosa¹, Bradley Voytek^1,2,3; ¹Halıcıoğlu Data Science Institute, UC San Diego, ²Neurosciences Graduate Program, UC San Diego, ³Department of Cognitive Science, UC San Diego

Assigning functional network labels (e.g., default mode, frontoparietal control, salience) to fMRI activation maps is common yet fragile: network topographies and names vary across atlases, and labeling often relies on visual heuristics and individual expertise, limiting reproducibility. To automate this process, we train a contrastive model with 30,000 neuroimaging publications, aligning latent text and neuroimage representations. Contrastive learning aligns matched text–map pairs in a shared embedding space while separating mismatches, enabling zero-shot network labeling using cosine similarity between a query map embedding and candidate label embeddings. This is compatible with any set of text, including terms or descriptions of cognitive processes, network names, or anatomical regions. The contrastive model may also be used to rank a set of publication text most similar to a query map. For external validation, we use group-level activation maps from NeuroVault to ensure titles and abstracts are well aligned with corresponding statistical maps and achieve area under the recall@k curve = 0.85 for ranking and retrieval. For network labeling, we evaluate against a collection of network atlases and obtain 80% accuracy. Finally, we demonstrate applicability to labeling ICA components from the Human Connectome Project and UK Biobank. These results demonstrate that contrastive neuroimage–language alignment offers a scalable, flexible, and literature-grounded method for automated network labeling tasks.

Topic Area: METHODS: Neuroimaging