A Deep Learning Model for Pancreatic Ductal Adenocarcinoma Chemotherapy Outcome Prediction

Abstract

Pancreatic Ductal Adenocarcinoma (PDAC) is an aggressive abdominal malignancy, with an overall 8.5% 5-year survival rate. PDAC is often detected too late for surgical resection and associated with resistance to chemotherapy and radiation. Morphological characteristics of PDAC tumors can be extracted from CT scans and are associated with tumor characteristics and behavior. In this research, a deep-learning system for predicting chemotherapy outcome based on CT scans is being explored. To establish the foundation for this system, a comparative analysis between 2D-UNet and 3D-UNet has been performed. Experiments reveal 2D-UNet model to have higher accuracy in pancreas segmentation. To increase prediction accuracy, the effects of novel data augmentation techniques, including window level and cropping, have been utilized. By establishing techniques for standardizing CT scans from different datasets and using a flexible segmentation model, we aim to create a pipeline for pancreas segmentation followed by tumor extraction and staging using texture analysis.