Improving Colorectal Cancer Screening and Risk Assessment through Deep Learning on Medical Images and Records

National Library of Medicine (NLM), National Institutes of Health (NIH)
NLM Express Research Grants in Biomedical Informatics, R01LM012837

02/12/2019 – 01/31/2023

Saeed Hassanpour, PhD (Geisel School of Medicine at Dartmouth College)

Arief Suriawinata, MD (Co-I); Lorenzo Torresani, PhD (Co-I); Lynn Butterly, MD (Co-I)

Most colorectal cancer cases start as a small growth, known as a polyp, on the lining of the colon or rectum. Although colorectal polyps are precursors to colorectal cancer, it takes several years for these polyps to potentially transform into cancer. If colorectal polyps are detected early, they can be removed before they can progress to cancer. The microscopic examination of stained tissue from colorectal polyps on glass slides—the practice of histopathology—is a key part of colorectal cancer screening and forms the current basis for prognosis and patient management. Histopathological characterization of polyps is an important principle for determining the risk of colorectal cancer and future rates of surveillance for patients; however, it is time-intensive, requires years of specialized training, and suffers from high variability and low accuracy. In addition, as is evident by the domain literature, other health factors, such as medical and family history, play an important role in colorectal cancer risk; however, they are not considered in current standard guidelines for colorectal cancer risk assessment. Therefore, there is a critical need for computational tools that can incorporate both histopathological and relevant clinical/familial information to help clinicians better characterize colorectal polyps and more accurately assess risk for colorectal cancer.

To address this critical need, this application proposes to build a novel, automatic, image-analysis method that can accurately detect and classify different types of colorectal polyps on whole-slide microscopic images. The proposed approach will be able to identify discriminative regions and features on these images for each colorectal polyp type, which will provide support and insight into the automatic detection of colorectal polyps on whole-slide images. Finally, this project will provide an accurate risk prediction model to integrate visual histology features from microscopic images with other risk factors and relevant clinical information from medical records for a comprehensive colorectal cancer risk assessment. The proposed image analysis and prediction methods in this project are based on a novel deep-learning methodology and rely on numerous levels of abstraction for data representation and analysis. The technology developed in this proposal will be rigorously validated on data from patients undergoing colorectal cancer screening at the investigators’ academic medical center and on the records from the New Hampshire statewide colonoscopy data registry. Upon successful completion of this project, the proposed bioinformatics approach is expected to reduce the cognitive burden on pathologists and improve their accuracy and efficiency in the histopathological characterization of colorectal polyps and in subsequent risk assessment and follow-up recommendations. As a result, this project can have a significant, positive impact on improving the efficacy of colorectal cancer screening programs, precision medicine, and public health.

Colorectal cancer is the second leading cause of cancer deaths in the United States; however, it can be prevented through regular screening. The proposed project is expected to reduce the manual burden and potential errors of diagnosis, risk assessment, and follow-up recommendations in colorectal cancer screening. Therefore, the outcomes of this project can potentially reduce screening time and costs, eliminate undue stress to patients, increase the coverage and accuracy of screening programs