Transformer-based deep learning model for integrated pathomics and radiomics in predicting postoperative survival of intrahepatic cholangiocarcinoma patients.

Abstract

e16328 Background: Intrahepatic cholangiocarcinoma (ICC) is a highly aggressive malignancy with a dismal prognosis. While surgical resection stands as the cornerstone of curative treatment, the absence of reliable predictive tools for postoperative outcomes poses a significant challenge in optimizing patient management. Methods: We developed a state-of-the-art multimodal transformer-based model, integrating patient-specific proteomic data to provide a comprehensive understanding of ICC in surgical patients. To ensure robustness, the model was rigorously validated using an independent external dataset. Results demonstrated its exceptional ability to predict postoperative survival in ICC patients, while also uncovering novel insights into the molecular and radiomic factors linked to poor prognosis. By leveraging these predictive outcomes, we further analyzed the proteomic dataset, revealing intricate associations between radiomic signatures and proteomic profiles, which may play critical roles in ICC progression and outcomes. Results: In this study, we developed a multimodal transformer-based model leveraging MRI data and pathology data (whole-slide images and tumor-cell nuclei) from a cohort of 301 patients. The model was validated in two independent datasets (n = 100 and n = 80), achieving AUCs of 0.839 and 0.852, respectively, with token-level AUCs of 0.853 and 0.874. Further validation in an external test cohort (n = 36) demonstrated robust performance, yielding an AUC of 0.867. Integrated analysis of proteomic and radiomic models revealed that the upregulation of CDX1 was significantly associated with prognosis, highlighting a group of highly interactive proteins that may play pivotal roles in the underlying biological processes of intrahepatic cholangiocarcinoma. Conclusions: The integration of multimodal omics models provides a powerful and reliable framework for predicting postoperative prognosis in ICC patients.