Min Yuan

Ph.D

Anhui Medical University

Professor

PLCO (Learn more about this study)

PLCOI-783

May 7, 2021

Deep Learning for Investigation of Relationship between Whole Slide Image Features, Biomarkers, and Clinical Outcomes in Cancer Patients

Digital pathology, particularly with the use of digital whole slide images (WSIs), has emerged as a novel tool in clinical studies and has potential to accurately and objectivity predict various clinical endpoints and patient prognosis, providing new insights on pathology and cancer biology, and allowing further risk stratification of patients and refinement of treatment decisions in clinic. Digital pathology is primary based on deep learning (eg, convolutional neural networks (CNN) and graph neural networks (GNN), etc). Courtiol and co-workers developed a deep convolutional neural network based on whole-slide digitized images to accurately predict the overall survival of mesothelioma patients without annotations from pathologists. Saillard et al. have further refined this model and demonstrated that artificial intelligence can help improve the prediction of the prognosis of hepatocellular carcinoma.

There is an urgent need to develop reliable pathology image-based tools or biomarkers in routine clinical practice to further assist identification of high-risk patients in cancer patients. In this study, we will utilize deep learning models to identify associations among whole-slide histological image features, peripheral/genetic/molecular biomarkers (genetic, lab tests, hormones, immune markers, etc) and clinical outcomes collected in the PLCO study. This proposed study can help us to better understand cancer biology and provide more accurate prediction of disease prognosis than conventional approaches.

The objective in this research will be to better understand cancer biology and to provide more accurate prediction of clinical outcomes (e.g., survival) and clinical characteristics (eg. Cancer staging) in cancer patients using H&E whole slide images and peripheral biomarkers. Specifically,

a. To construct convolutional neural network (CNN or/and GNN) based deep learning models to extract features from H&E image slides
b. To train a deep-learning neural network (DNN) or/and statistical prediction models to predict clinical outcomes and clinical characteristics for cancer patients
c. To train a deep-learning neural network (DNN) or/and statistical prediction models to investigate associations between image features and peripheral/genetic/molecular biomarkers (genetic, lab tests, hormones, immune markers, etc)
d. To identify novel imaging features and biomarkers that are of predictive values for clinical outcomes and clinical characteristics for 4 cancer types in PLCO
e. To construct multivariate models using identified image features and biomarkers for predicting clinical outcomes and clinical characteristics