Accurate Detection of Early-Stage Pancreatic Tumors using Artificial Intelligence Analysis of Abdominal Computed Tomography Scans

Project Summary Accurate detection of Pancreatic Ductal Adenocarcinoma (PDAC) is highly challenging, especially in its early stages, given the lack of specific symptoms and diagnostic biomarkers. The current reliance on abdominal computed tomography (CT) for PDAC screening has limitations, making it complicated to identify small pancreatic tumors in the early stages, contributing to over 80% of late-stage diagnoses. Late-stage diagnosis results in limited surgical options, with only a 15% resection rate and a 5-year overall survival rate of 13%, which could potentially increase to 50% with early detection. Previous studies have emphasized the variation in pancreatic tumor histology across pancreatic subregions (PSRs) and changes in pancreas morphology associated with PDAC incidence due to underlying complications associated with PDAC and changes in the genetic and molecular profiles. We propose a comprehensive examination of cancerous pancreas in diagnostic CT scans and the development of an AI tumor detection model for PDAC in CT images, with a focus on small pancreatic tumors. The model will undergo three major enhancements: analyzing CT texture variations in head, body, and tail tumors, integrating distinctive morphological features of the pancreas, its subregions, peri-pancreatic fat, and liver that indicate the presence of cancer, and incorporating non-imaging factors associated with PDAC. Firstly, An AI model will be developed that will utilize this tumor heterogeneity and perform a more targeted detection of tumors. Secondly, the detection model will be enhanced by integrating the distinctive morphological features of pancreas, PSRs, peri-pancreatic fat and liver that are indicative of PDAC and can potentially be used as secondary signs of cancer. Using these features, a machine learning model will estimate a probability of presence of PDAC, which will assist accurate interpretation of detected tumors. Thirdly, the enhanced model will be further improved by incorporating the non-imaging factors (e.g., clinical- and blood-based biomarkers) that are often associated with PDAC. Each phase of the project will utilize a large dataset of abdominal CT scans (1064 healthy and 1064 diagnostic). The model will undergo thorough evaluations at each stage, aiming for incremental accuracy improvements. The ultimate goal is to achieve a detection accuracy for small pancreatic tumors, currently around 70%, to be elevated to at least 90%. The project's endpoint is a rigorously trained and validated AI model for PDAC detection, intending to support follow-up screenings for high-risk individuals or those displaying early symptoms. This approach aims to minimize the risk of misdetection and misinterpretation of early or late-stage PDAC tumors, facilitating early intervention in the treatment of this potentially fatal disease.