Foundation-model imaging biomarkers for pancreatic cancer: cross-species discovery of conserved tumour biology in mouse MRI and human CT

BACKGROUND: Computed tomography (CT) stages pancreatic ductal adenocarcinoma (PDAC) but does not read out the treatment-guiding tumour biology, requiring biopsy. Whether an imaging-biology relationship discovered in preclinical MRI holds independently in human CT is unknown. METHODS: One pipeline was applied to two PDAC cohorts: longitudinal 9.4 T multi-parametric MRI (169 sessions; 1002 image series, 471 quantitative maps) with terminal histopathology in male C57BL/6 mice bearing orthotopic KPC-derived tumours treated with gemcitabine, a dendritic-cell vaccine, both, or neither, and abdominal CT with matched proteogenomics from CPTAC-PDA. Radiomics and frozen foundation-model embeddings were modelled under grouped cross-validation against five pre-specified paired endpoints, with no cross-species transfer of features, weights, or labels. FINDINGS: Applied independently, the pipeline recovered four matched molecular endpoints in both species (CD8 infiltration, stromal collagen, CD8:CD68 immune balance, cytolytic activity), with areas under the receiver-operating-characteristic curve (AUROCs) of 0.938-0.960 in mice and 0.780-0.856 in humans. Survival concordance indices were 0.729 and 0.689. After adjustment for age, stage, and grade, the human imaging score remained independently prognostic (hazard ratio 1.87 per SD, 95% CI 1.32 to 2.66; p < 0.001 [Wald test]; combined concordance index 0.726 versus 0.611 for clinical covariates). Findings held in portal-venous-restricted analysis; frozen embeddings outperformed end-to-end fine-tuning in both arms. INTERPRETATION: One pipeline recovered matched PDAC biology in mouse MRI and human CT, supporting preclinical imaging to nominate non-invasive CT biomarker candidates for biopsy-sparing PDAC trial stratification, pending prospective external validation.