Skip to Main Content

COVID-19

What people with cancer should know: https://www.cancer.gov/coronavirus

Get the latest public health information from CDC: https://www.cdc.gov/coronavirus/

Get the latest research information from NIH: https://covid19.nih.gov/

Principal Investigator
Name
Okyaz Eminaga
Degrees
Ph.D/MD
Institution
Stanford University School of Medicine
Position Title
AI-based prognostic scoring system for prostate cancer
Email
About this CDAS Project
Study
PLCO (Learn more about this study)
Project ID
PLCOI-635
Initial CDAS Request Approval
Jun 18, 2020
Title
AI-based scoring system for prostate cancer
Summary
Prostate Cancer (PCa) is known to exhibit diverse clinical behaviors, ranging from indolent to lethal diseases. A
key clinical need is to identify characteristics that distinguish indolent from advanced disease to direct treatment
to the latter. In this context, the Gleason grade group (GG) is a malignancy PCa grading system that helps
determine advanced diseases. Although GG is a robust prognostic parameter, 20-40% of patients are found to
suffer from occult higher-grade cancer at prostatectomy (surgical removal of the prostate to treat the cancer) in
comparison to the original GG found in the biopsy. This discrepancy has the greatest implications for patients
with presumed low-grade cancer, who are currently advised to forgo treatment and simply watch their cancers
(Active Surveillance). Failure to diagnose higher-grade cancer in these men could lead to progression of occult
high-grade cancer. Our long-term goal is to improve the PCa management by providing clinically meaningful
decision-aided tools. Here, the overall objective, which denotes the next step in pursuit of that goal, is to
determine whether or not the features extracted by deep learning, which is part of artificial intelligence (AI), from
histology images of biopsy samples are more reliable in determining high-grade PCa than the current Gleason
grade groups. From the Genome Cancer Atlas dataset, our result implies that PlexusNet, a novel neural network
architecture, facilitates the extraction of independent prognostic and explainable feature scores from histology
images. Further, these feature scores are correlated with genomic alterations as well as molecular subtypes.
The central hypothesis is that having features determined by AI approaches from the histology images are found
to be consistent between biopsy samples and RPE and accurately detect high-grade PCa. In order to test this
hypothesis, the proposed project will extensively evaluate the potential of biologic and prognostic feature scores
determined by AI based on an analysis of histology images taken at biopsy for identifying high-grade PCa and
reducing false-negative cases. The objective of the proposed research is as follows: once it is known that AI can
lower the risk of missing high-grade PCa at the biopsy, it becomes possible to develop computer-aided tools that
justify the diagnosis of advanced PCa at biopsy. We plan to test our central hypothesis and, thereby, accomplish
our overall objective by pursuing the following three specific aims that plan (Aim 1) to identify
features for advanced PCa from histology images; (Aim 2) to integrate the feature scores into a risk prediction
model for high-grade PCa; (Aim 3) to optimize the current clinical decision algorithms. Upon conclusion, we expect to introduce an AI-based risk prediction system for high-grade PCa. The results are likely to contribute to developing a comprehensive scoring system for PCa so
as to improve the identification of cases with high-grade PCa. Further, the results are expected to improve the
patient selection for active surveillance by reducing the false-negative rates for high-grade PCa.
Aims

Aim 1: To identify features for advanced PCa from histology images
We will utilize deep learning models to extract histopathological feature scores from pathology images more prognostic or predictive than Gleason grading. Patch images of segmented lesions will be generated to evaluate the prognostic feature scores from histology images of PCLO.
Aim 2: Integration of feature scores into a risk prediction model for aggressive PCa risk.
A machine learning model for the risk prediction of high-grade PCa, including the feature scores and the contemporary prostate biopsy risk score will be developed and evaluated in order to reduce false-negative rates on cases covering different treatment conditions of PCa and GG upgrading.

Aim 3: Optimization of the current clinical decision algorithms using AI
Here, we will consider cases that have complete longitudinal data to evaluate whether AI can improve clinical decision making from diagnosis to follow-up.

Collaborators

Mahmoud Abbas, Institut für Pathologie und Zytologie
Axel Semjonow, Prostate Center, University Hospital Muenster