Identifying Survival Phenotypes in High-Grade Serous Ovarian Cancer by Application of AI to Digital Pathology
Principal Investigator
Name
Ernst Lengyel
Degrees
MD PhD
Institution
University of Chicago
Position Title
Chair
Email
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCOI-1737
Initial CDAS Request Approval
Nov 12, 2024
Title
Identifying Survival Phenotypes in High-Grade Serous Ovarian Cancer by Application of AI to Digital Pathology
Summary
Ovarian cancer is a lethal gynecologic cancer with a median survival less than 5 years, however approximately 10% of patients prove themselves long-term survivors (LTS) and live more than 10 years after initial diagnosis despite similarly aggressive initial presentation. Prior attempts to identify predictive features clinically and genomically that differentiate LTS from those following the typical survival curve of ovarian cancer have found correlations but none of the identified features have proven consistently predictive. We aim to use state-of-the-art machine learning methods to predict survival of patients with ovarian cancer using H&E whole-slide images (WSI). Further, we plan to use the metadata generated by the prediction to investigate the histologic differences found at the extremes of survival. The ovarian cancer population from PLCO will serve as an external validation set for our models.
Our preliminary models, trained on an internal cohort of approximately 90 patients with high-grade serous ovarian cancer, attain an AUROC between 0.65-0.73 in predicting the extremes of survival classification (Short-term survival - STS: < 2 years; LTS: > 7 years). Initial technical modifications have thus far yielded incremental improvements of 0.05-0.15 in the AUROC per iteration on our internal test set. With each improvement in performance, we have noted that model attention narrows; we plan additional pathologist-led analyses to identify trends in these regions of attention.
Provided we are successful in predicting survival from H&E WSIs and able to effectively validate on external datasets such as PLCO, we plan to develop a web-based interface to establish our models as a tool in clinical practice, providing patients with real-time prognostic estimations.
Our preliminary models, trained on an internal cohort of approximately 90 patients with high-grade serous ovarian cancer, attain an AUROC between 0.65-0.73 in predicting the extremes of survival classification (Short-term survival - STS: < 2 years; LTS: > 7 years). Initial technical modifications have thus far yielded incremental improvements of 0.05-0.15 in the AUROC per iteration on our internal test set. With each improvement in performance, we have noted that model attention narrows; we plan additional pathologist-led analyses to identify trends in these regions of attention.
Provided we are successful in predicting survival from H&E WSIs and able to effectively validate on external datasets such as PLCO, we plan to develop a web-based interface to establish our models as a tool in clinical practice, providing patients with real-time prognostic estimations.
Aims
1. To determine the optimal machine learning architecture to classify H&E WSIs of ovarian cancer into survival phenotypes
2. To identify and characterize the regions with H&E WSIs of ovarian cancer that are most predictive of survival
Collaborators
Ernst Lengyel, MD PhD, University of Chicago
Leonhard Donle, PhD, University of Chicago
Hilary Kenny, PhD, University of Chicago
Alex Pearson, MD, University of Chicago
Emma Rose Carpenter, University of Chicago
Elena Nevins, University of Chicago