Corey Arnold

Ph.D.

University of California, Los Angeles

Professor

PLCO (Learn more about this study)

PLCOI-1492

Mar 11, 2024

Computational Feature Profiling and Modeling for Prostate Cancer Detection and Risk Stratification

Prostate cancer is the most common and second deadliest non-skin cancer in American men, accounting for 26% of new cancer diagnoses and 9% of cancer deaths in men. Active surveillance, radical prostatectomy and radiotherapy are commonly used treatments for clinically localized prostate cancer. However, current risk stratification methods cannot be used effectively to avoid subjecting patients with clinically indolent cancers to unnecessary interventions, causing significant morbidity and cost. The primary components currently involved in screening are the digital rectal exam (DRE) and serum biomarkers, such as PSA, PCA3, PHI, and 4Kscore. Unfortunately, despite advances in these tests, overdiagnosis remains a major problem due to limited specificity. As a result, 90% of patients diagnosed with prostate cancer receive treatment, even though up to 60% of those patients could be candidates for active surveillance. Such treatment often results in long-term reductions in functional outcomes.

The research objective of this project is to develop novel markers and models to both more accurately detect aggressive cancer and to forecast its arrival. Using a large cohort of patients, we first plan to identify novel pathomic and germline features that indicate the presence of aggressive cancer or its precursors. We then plan to implement an integrative graph convolutional network (GCN) combined with a convolutional neural network (CNN) to generate new multi-modal representations of underlying cancer state within the entire prostate. The framework will combine multiparametric magnetic resonance imaging (mpMRI), digital histology images, germline features, biomarkers, and other predictors. We will also implement a baseline nomogram risk model for comparison, as well as several new nomogram models that incorporate our newly identified features.

Using prostatectomy specimens, learn multi-resolution pathomic phenotypes from low-grade and benign tissue regions that indicate the presence of aggressive cancer, and confirm the replicability of the phenotypes in biopsy specimens.

Develop models that use interpretable pathomic features to predict clinical outcomes.

Develop a combination convolutional neural network-graph convolutional network (CNN-GCN) that can process histology data from either biopsies or prostatectomies and fuse it with imaging features.

Extend the CNN-GCN to include an additional module that incorporates patient-level features.

Construct new nomogram models that include known risk factors as well as new imaging, histology, and germline features and compare their performance to the CNN-GCN and to an established nomogram.