Survival prediction in pancreatic cancer by attention-driven feature extraction on histopathology whole slide images: a multi-cohort validation

Pancreatic ductal adenocarcinoma (PDAC) is an aggressive disease with a dismal prognosis. Despite efforts to improve therapy outcomes in PDAC, overall survival remains at 2 to 5 years following initial diagnosis. To date, there are no established predictive or prognostic biomarkers for PDAC tumors. The availability of digitized H&E stained whole slide images (WSI) has led to an uptake in deep learning-based approaches toward comprehensive, automatic interrogation of tumor-specific attributes for disease diagnosis and prognosis. However, a significant challenge with the interrogation of large WSIs (gigabytes in size) is that only a small portion of the tissue (i.e. ROIs) contains information pertinent to diagnosis or prognosis. In this work, we investigated whether "highattention" ROIs (i.e. patch regions) identified by an attention-driven model to differentiate tumor from benign regions, may also be associated with survival outcomes in PDAC patients. The attention model was developed using a total of n = 461 WSI of H&E-stained pancreatic tumors, from two public repositories. Our approach first identifies attention maps (i.e. ROIs) using clustering-constrained-attention multiple-instance learning (CLAM), on WSI labeled as PDAC versus benign pancreas. Subsequently, the learned attention maps are employed within a LASSO regularized Cox-hazard proportional model to distinguish between high and low survival-risk groups of PDAC patients. Results were evaluated via a log-rank test and compared with established demographic variables (age, sex, race) to predict survival risk. While individual demographic variables did not demonstrate significant differences in survival risk, the attention-driven WSI features yielded significant stratification of low and highrisk groups in both the training (p = 0.0014, Hazard Ratio (HR), 2.0 (95 % Confidence Interval (CI) 1.3 -3.1)) and the test set (p = 0.0012 HR = 2.0 (95 % CI 1.3 -2.6)). Following a large, multi-institutional validation, our deep-learning approach may allow for designing more precise prognostic and predictive histopathological biomarkers for PDAC tumors.