Explainable Multimodal Prognostic Modeling Uncovers New Clinically Relevant Pathological Features in a Multi-Institutional Pancreatic Neuroendocrine Tumor Cohort

A multimodal continuous-time survival model for surgically resected pancreatic neuroendocrine tumors links H&E image attention to prognosis, with late-fusion C-indices up to 0.88.

• Why it matters: Existing predictors like Ki-67 and stage incompletely explain why similar PanNET patients have different outcomes, and it is unclear what additional pathology signals in routine slides matter.
• What they did: The study built Cox models from structured clinicopathologic data and deep-learning models from ~30,000 histology images in 1,438 patients, then used late multimodal fusion and attention-based interpretation.
• The result: The clinical model reached ~0.87 C-index, image-only ~0.80, and late fusion ~0.88 internally; attention highlighted tumor leading edges and vascular/perineural invasion, with stroma comprising nearly 50% of high-attention regions.

AI summary of the abstract · may contain errors · not medical advice · read the paper.