Deep learning-based multimodal approach for non-invasive prediction and prognostic analysis of immune and angiogenic biomarkers in extrahepatic cholangiocarcinoma.

Objective: This study aims to develop a deep learning (DL)-based multimodal framework that integrates magnetic resonance imaging (MRI), clinical, and laboratory data to predict programmed death ligand 1 (PD-L1) and vascular endothelial growth factor (VEGF) expression in Extrahepatic cholangiocarcinoma (eCCA) patients and assess the prognostic value. Methods: A retrospective cohort study involving 96 patients with eCCA was conducted across two institutions. A total of 16050 raw MRI images (11505 T1WI, 2371 T2WI, 2372 DWI) and 1570 tumor-containing images (990 T1WI, 289 T2WI, 291 DWI) were analyzed. Radiomic feature extraction was performed manually segmented tumor regions from MRI scans. The multimodal DL framework integrated DL features extracted from images and radiomic features as well as clinical-laboratory features through a repeated attention mechanism. Prognostic stratification was performed using Cox regression analysis to predict overall survival (OS) and evaluate the clinical utility of the model. Results: The DL framework demonstrated moderate predictive performance for PD-L1 expression (AUC = 0.71) and good predictive capability for VEGF expression (AUC = 0.85) in the test cohort. The combination of DL-based imaging features and radiomic data outperformed single-modality approaches. Prognostic analysis revealed significant associations of model-predicted PD-L1 and VEGF expression with OS in eCCA patients. The Cox model-based nomogram demonstrated significant survival stratification (p = 0.006), with performance comparable to traditional immunohistochemistry-based methods. Conclusion: Our findings highlighted the potential of integrating DL and radiomics for non-invasive, preoperative biomarker profiling, offering a promising tool for personalized treatment strategies and improved clinical decision-making in eCCA.