Skip to Main Content

An official website of the United States government

Discovery and validation of computational morphologic features to predict prostate cancer recurrence (v2)

Principal Investigator
Name
Michael Cook
Degrees
PhD
Institution
NCI
Position Title
Investigator
Email
About this CDAS Project
Study
PLCO (Learn more about this study)
Project ID
PLCO-187
Initial CDAS Request Approval
Jan 8, 2016
Title
Discovery and validation of computational morphologic features to predict prostate cancer recurrence (v2)
Summary
If prediction algorithms for prostate cancer recurrence can be improved, surveillance and treatment decisions could be refined which would benefit patient outcomes. Current algorithms have an area under the receiving operating characteristic curve of ~0.70. Gleason score is an essential component of such algorithms, yet the burgeoning field of computational morphologic analysis may be able to provide cheap and robust predictive features that complement the Gleason metric and improve prediction. Therefore, we propose to use state-of-the-art computational morphologic analysis on H&E prostate cancer tissue microarrays available in the Prostate Cancer Progression study nested within the PLCO Cancer Screening Trial. This study provides tumor and adjacent normal TMA cores for each of 1003 men who were diagnosed with localized prostate cancer and who underwent radical prostatectomy. For each TMA core, we will use automated, computational scoring of a rich, quantitative feature set, including measures of size, shape, intensity, texture, and variability. Machine learning will then assess whether features are predictive of prostate cancer recurrence. The best morphologic features will be added to existing algorithms to assess whether they can improve prediction of prostate cancer recurrence.
Aims

The overall goal of the study is to gain insight on whether computational morphologic features from H&E prostate cancer and adjacent normal core TMAs can aid prediction of prostate cancer recurrence.

The specific aims of the study are:

1. To discover and internally validate computational morphologic features associated with prostate cancer recurrence.

2. To assess whether validated morphologic features significantly increase the predictive accuracy of the CAPRA score.

3. To assess whether an AI-powered Gleason grade tissue segmentation increases the consistency and accuracy of pathologist-rated Gleason score.

Prostate cancer is the leading form of malignant disease in American males. Unfortunately, current screening and monitoring methods for prostate cancer suffer from significant error rates, namely high false positive rates. This mischaracterization stems from the heavy reliance on Gleason scoring of patterns of cell morphology and tissue architecture, which yields complex results (scores of 1-5 for primary and secondary patterns, creating a patient’s Gleason Score of 2-10) and high error rates. Thus, there is a strong need for a consistent, accurate, and standardized scoring method for grading prostate cancers.

In order to meet this need for an improved method of diagnosing and characterizing prostate tissue samples, PathAI is developing an Artificial Intelligence (AI)-based approach that applies computer vision and deep learning analysis with routine hematoxylin and eosin (H&E) stained tumors to discriminate aggressive prostate malignancies from indolent tumors based on architectural, color, and nuclear patterns. PathAI will apply its platform to generate an AI-powered Gleason grade tissue segmentation to assist pathologist decision making regarding whether a patient is at high-risk for aggressively invasive or malignant disease.

In order to determine whether these predictions increase the consistency and accuracy of pathologist assessment from routine histology images, we propose leveraging the PLCO Prostate Cancer Progression cohort for Gleason scoring tasks both with and without the aid of an AI-assisted evaluation. In this experiment, we will ask a panel of pathologists to assign Gleason scores to a set of tissue samples and assess whether the availability of an AI-powered image segmentation increases the inter-pathologist concordance on these routine grading tasks. After a given “wash out” period, we will reassign images to the pathologist cohorts for their evaluation with and without the image segmentation overlays. All experiments will be performed on the PathAI platform using board-certified pathologists with prior experience in signing out prostate cancer cases. The result of this experiment will allow to assess whether inter-observer variability can be mitigated by the aid of an AI-powered decision support tool.

Collaborators

Andrew H Beck, Harvard Medical School
Amanda Black, NCI, DCEG
Hormuzd Katki, NCI, DCEG
Humayun Irshad, Harvard Medical School
Jong Cheol Jeong, Harvard Medical School
Larry Sternberg, NCI-Frederick
Robert Hoover, NCI, DCEG
Sarah Daugherty, PCORI
Sindhu Ghanta, Harvard Medical School