Secondary Logo

Journal Logo

Informatics Technologies to Objectively Stratify Prostate Cancer Risk

Varghese, Bino, PhD

doi: 10.1097/01.COT.0000557849.83974.9d
News
Free

Prostate cancer (PCa) is the most prevalent male malignancy worldwide. Treatment recommendations are currently based on risk stratification using a variety of factors including but not limited to PSA, Gleason score, and T category. A recent large study with data from 17,943 patients on active surveillance (low-risk PCa patients) treated with radical prostatectomy showed that upgrading, upstaging, or nodal metastases occurred in 45 percent of men (Urol Oncol 2015;33:164.e11-17). Clearly, we need better tools to accurately assess risk stratification and optimize effective use of active surveillance. The ability to accurately assess the risk of a diagnosed PCa tumor could also improve the selection of appropriate treatment for these patients, leading to improved outcomes, including PCa-specific mortality.

Over the past decade, multi-parametric MRI (mpMRI) has become increasingly important for the evaluation, localization, and staging of PCa (Urol Oncol 2015;33:337.e15-24, Eur Urol 2015;68:1045-1053). In combination with the Prostate Imaging Reporting and Data System Version 2 (PI-RADS v2), encouraging results have been reported for the prediction of the likelihood of intermediate- and high-grade cancers (Acad Radiol 2017;24:1101-1106, J Magn Reson Imaging 2017;45:579-585).

However, despite the high sensitivity of mpMRI, assessing PCa is still a visual, and therefore subjective, process (J Magn Reson Imaging 2017;45:579-585). The reported inter-observer agreement has only been moderate to good (Acad Radiol 2017;24:1101-1106, J Magn Reson Imaging 2017;45:579-585), with several multi-reader studies finding an overall inter-reader agreement ranging from poor (0.5) to reasonable (0.71), depending on the study and reader experience (Acad Radiol 2017;24:1101-1106, J Magn Reson Imaging 2017;45:579-585, Radiology 2015;275:458-468, J Urol 2016;195:1428-1435).

The discordance between patients' inferred clinical risk and their imaging findings makes it difficult to interpret tumor imaging data. Extracting objective features from these images (radiomics) and subsequently using these features to develop frameworks based on machine learning (ML) methods can augment radiologists' role in clinical care by providing more objective risk assessments.

Radiomics deals with the extraction of quantitative features such as texture, size, and shape from routine clinical images (Radiology 2015;278:563-577, Nat Commun 2014;5:4006). The underlying assumption is that images acquired during routine clinical care contain latent information regarding tumor behavior that can be quantified using a variety of image characterization algorithms (Radiology 2015;278:563-577). Extracting these radiomic metrics enables the conversion of collections of digital clinical images into structured, quantitative data that can help model tumor behavior. For example, in radiomic studies conducted using mpMRI, a variety of texture metrics aided in the diagnosis of PCa (Med Phys 2011;38:83-95, J Magn Reson Imaging 2009;30:161-168, Eur Radiol 2015;25:2840-2850).

By design, ML methods sift through large amounts of high-dimensional data without any particular guiding (biomedical) hypothesis to directly discover potentially actionable knowledge (Introduction to Machine Learning. MIT Press; 2014, Machine Learning in Medicine-a Complete Overview. Springer; 2015). These abilities make ML methods, especially those for classification, ideal for radiomic studies that aim to improve PCa assessment and make it less subjective (Abdom Radiol (NY) 2018; https://doi.org/10.1007/s00261-018-1660-7, Transl Cancer Res 2016;5:432-447).

Previous work combining radiomics and ML for PCa assessment has been limited to the use of one or a small number of ML algorithms used specifically for classification, methods, and their evaluation using standard measures like AUC. Developing an automated framework that systematically and rigorously identifies the best classifier(s) for predicting tumor risk with a given set of radiomic features can substantially boost the potential of this approach. We developed such a framework comprised of classification, cross-validation, and statistical analyses to identify a classifier that most accurately differentiates high-risk PCa patients from lower-risk ones in a sizeable cohort examined using mpMRI.

Our final framework examined seven common classifiers known for their performance and stability for predicting PCa risk, using 110 radiomics features extracted from T2- and diffusion-weighted mpMRI images. Our cohort consisted of 121 PCa patients that were randomly split into training (n=68) and validation (n=53) sets prior to the application of our framework. Using a systematic cross-validation setup and rigorous statistical analysis of the performance of candidate classifiers evaluated on the training set, the framework identified the quadratic kernel support vector machine (QSVM) classifier as the most effective PCa risk prediction method.

Indeed, the QSVM classifier performed well on the independent validation set across a variety of evaluation measures (AUC=0.71, F-measure=0.69, Precision=0.57, and Recall=0.86). In particular, it performed better than PI-RADS v2, the imaging score-based method currently used to assess PCa risk, especially in terms of class-specific evaluation measures, namely F-measure (0.52), Precision (0.45), and Recall (0.61).

The use of these class-specific performance metrics, also a novelty of our framework, enables the assessment of a classifier's performance on perhaps the most difficult category of patients, such as the high-risk PCa patients in our cohort. Our results demonstrate the effectiveness of data-driven frameworks for assessing and deriving objective imaging-based risk predictors that can assist radiologists in improving clinical care.

In the future, we propose to further evolve our pipeline and workflow for a more comprehensive evaluation in patients with other malignancies.

BINO VARGHESE, PHD, is Assistant Professor of Research Radiology in the Keck School of Medicine of USC, Los Angeles.

This project was a collaborative endeavor between the Radiomics lab led by Vinay Duddalwar, MD, at University of Southern California (USC) in the Department of Radiology; the USC Institute of Urology at USC led by Inderbir Gill, MD; and the Pandey Lab led by Gaurav Pandey, PhD, at the Icahn School of Medicine at Mount Sinai. The Radiomics Lab is a research group in the Radiology Department at USC with a multidisciplinary team of radiologists, engineers, programmers, and statisticians who are committed to developing image analysis tools, cutting-edge clinical workflows from standard of care imaging and personalized precision medicine. Funding support from the Big Data to Knowledge (BD2K) Training Coordinating Center through the Data Science Rotations for Advancing Discovery (RoAD-Trip) program (grant #1U24ES026465-03) awarded to Bino Varghese, PhD, an Assistant Professor of Research working within the USC Radiomics Lab, aided in bringing together this multidisciplinary multi-institutional team to successfully complete this project.

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.
Home  Clinical Resource Center
Current Issue       Search OT
Archives Get OT Enews
Blogs Email us!