Deep Learning Pathomics for Predicting Clinical Outcomes in Pancreatic Ductal Adenocarcinoma

Abstract

Pancreatic ductal adenocarcinoma (PDAC) is among the most lethal cancers, with a five-year survival rate around 10%. Traditional prognostic methods, including imaging and blood markers like CA19-9, often fail to accurately predict therapeutic responses and outcomes. This research aims to enhance the prediction of disease-free survival (DFS) and overall survival (OS) in PDAC patients by developing a computational pathomics platform utilizing deep learning techniques.Whole slide images (WSIs) from The Cancer Genome Atlas (TCGA) and Fox Chase Cancer Center were segmented into patches of 224×224 pixels. A convolutional neural network (CNN) classification model was trained on annotated data to differentiate between cancerous and non-cancerous tissue patches, generating confidence scores for each classification. These scores were compiled into histograms reflecting the distribution of confidence levels across a patient's WSI.The histograms were integrated with clinical data to create a comprehensive dataset. Using the PyCaret machine learning framework, predictive models were developed to forecast DFS and OS based on combined pathomic and clinical features. The models demonstrated improved accuracy over traditional methods.By merging deep learning-based image analysis with clinical parameters, this pathomics approach provides a non-invasive, scalable method for enhancing prognostic assessments in PDAC. The study highlights the potential of integrating computational imaging and machine learning into clinical decision support systems. This work contributes to industrial and systems engineering by demonstrating how data analytics can address critical healthcare challenges, aiming to improve patient outcomes in pancreatic cancer management.