Deep learning on routine histopathology to enable one-year survival prediction in pancreatic ductal adenocarcinoma.

Abstract

e16016 Background: Pancreatic cancer ductal adenocarcinoma (PDAC) is an aggressive malignancy with a poor prognosis and limited tools for survival stratification. Our prior work demonstrated the feasibility of predicting Moffitt molecular subtypes from histopathology; however, these subtypes showed limited association with overall survival, motivating direct outcome-based modeling. We hypothesized that whole slide images (WSI) contain prognostically relevant morphologic signals that can be leveraged to predict short-term survival in PDAC. Methods: We analyzed PDAC cases from the Pancreatic Cancer Action Network (PANCAN) dataset (n = 846), including 272 patients with survival data. A WSI-based deep learning framework (PathRosetta) was trained to predict 1-year survival using five-fold cross-validation. In parallel, pretrained PathRosetta models predicting pan-cancer genetic alterations (APC, BRAF, KMT2D, KRAS, PIK3CA, TP53, and TTN) were used to extract mutation-relevant morphologic features for survival modeling, evaluated using leave-one-out cross-validation. The mutations were selected based on the availability of robust pretrained models with adequate case numbers in The Cancer Genome Atlas (TCGA), rather than PDAC-specific biological hypotheses. These represent mutations for which reliable pan-cancer prediction models currently exist. Results: At 1 year, 71 of 272 patients had died. Mutation-relevant morphologic-feature–based models demonstrated discrimination comparable to or superior to that of direct WSI modeling, with TP53 achieving the highest performance (Table). Conclusions: Deep learning applied to routine histopathology enables prediction of 1-year survival in PDAC. Leveraging mutation-relevant morphologic representations improves prognostic performance and supports a biologically grounded link between tumor genetics, tissue architecture, and clinical outcomes, highlighting the potential of WSI-based models for clinically meaningful risk stratification. It potentially aids in baseline risk stratification, with potential implications for clinical decision-making and trial enrichment. Performance for one-year survival Prediction in PDAC. AUC Accuracy Sensitivity Specificity PathRosetta (scratch) 0.768 ± 0.036 0.707 ± 0.018 0.634 ± 0.112 0.730 ± 0.054 APC 0.818 0.920 0.864 0.937 BRAF 0.791 0.905 0.841 0.926 KMT2D 0.786 0.904 0.851 0.922 KRAS 0.726 0.871 0.797 0.896 PIK3CA 0.729 0.890 0.771 0.933 TP53 0.877 0.935 0.938 0.934 TTN 0.745 0.862 0.846 0.868