Multimodal AI model superior to standard for prediction of key prostate cancer endpoints
SAN FRANCISCO — Prognostic biomarkers trained and validated through multimodal deep learning identified clinically relevant outcomes of men with localized prostate cancer better than standard clinical and pathologic variables, according to study results.
The findings, presented at ASCO Genitourinary Cancers Symposium, showed the feasibility of the massively scalable technology, which can help personalize prostate cancer management.
“We think AI tools represent a very good solution for global expansion of very useful prognostic biomarkers for prostate cancer,” Andre Esteva, PhD, head of medical AI at Salesforce Research, said during a presentation.
The current reliance on nonspecific tools such as PSA and digital rectal exam for prognostication in localized prostate cancer results in overtreatment and undertreatment of men, according to researchers. Other tools, such as Gleason score, can be inconsistent, and tissue assays have a long turnaround time, Esteva said. In addition, most tissue-based molecular biomarkers lack validation in prospective randomized trials.
Advantages of artificial intelligence tools using digital histopathology include performance across different needs, quick turnaround, robustness and widespread adoption, with no consumption of tumor tissue, he said.
Esteva and colleagues sought to develop and validate prognostic biomarkers using a novel multimodal artificial intelligence system (ArteraAI) with digital histopathology and clinical data from five randomized phase 3 NRG Oncology trials for prostate cancer. The researchers gathered histopathology image data of 5,654 men (median age, 70 years), obtaining 16.1 terabytes of data from 16,204 slides of pretreatment biopsy samples. Median follow-up of the studies was 11.4 years.
The investigators randomly divided trials into training (80%) and validation (20%) cohorts and compared the models with the NCCN model.
Using data from the slides, which Esteva stressed did not have annotations, they created an image quilt and divided it into patches that they used to train a self-supervised learning model. The clinical data and digital histopathology pipelines were combined into a third pipeline, which went through additional neural network layers to create an AI score developed to predict six clinical outcome variables. These included biochemical recurrence, distant metastasis, prostate cancer-specific survival and OS.
Evaluation of the trained and locked models on the validation cohort showed the MMAI prognostic model had better discrimination than the National Comprehensive Cancer Network model — which is based on clinical and pathologic features such as PSA, T stage, and Gleason score — for 5-year distant metastases (AUC = 0.83 vs. 0.73), 5-year biochemical recurrence (AUC = 0.67 vs. 0.58), 10-year prostate cancer-specific survival (AUC = 0.76 vs. 0.67) and 10-year OS (AUC = 0.65 vs. 0.58). The MMAI model also demonstrated superior performance vs. NCCN risk groups within each individual trial in the validation cohort for all clinical endpoints.
“Our ArteraAI prognostic models were successfully developed to predict long-term, clinically relevant outcomes for patients with prostate cancer,” Esteva said. “We believe we have so much power in this algorithm that we will continue to develop and improve over the next few years.”
Perspective
Back to Top
Current tools for predicting prostate cancer prognosis, such as Gleason score, PSA and tumor stage, are relatively nonspecific and, thus, have limited utility in tailoring therapy. In recent years, there has been an explosion in interest in applying AI tools to digitized histopathology whole-slide images. So-called deep learning (DL) models, in particular, have made remarkable advances in image recognition and are capable of identifying prognostically significant histomorphologic features that may not be apparent to a human viewer (eg, a pathologist).
In this study, the DL model developed and trained by Esteva and colleagues performed significantly better than a standard clinicopathologic prediction model (the NCCN model) in predicting all clinical outcomes.
These results are exciting and emphasize the power of DL models as tools for extracting novel prognostic features from digitized histopathology images. Unlike many other such studies, the model was validated in a large cohort of patients recruited from multiple clinical trials. This is important because DL models often produce poor results if they have not been trained on a sufficiently diverse (or a sufficiently large) data set.
Novel AI-based tools for prognostication are likely to become increasingly common and will no doubt be clinically deployed in the near future. Provided they are trained on large, diverse data sets, these tools offer the promise of more accurate prognostication, which will facilitate more targeted therapy and better patient outcomes.
Collapse
Esteva A, et al. Abstract 222. Presented at: ASCO Genitourinary Cancers Symposium; Feb. 17-19, 2022; San Francisco.
Read more about