Presentation Authors: Alex G. Raman*, Karthik V. Sarma, Los Angeles, CA, Wen Shi, Xi'an, China, People's Republic of, Steve Zhou, Alan M. Priester, Shyam Natarajan, William Speier, Steven S. Raman, Leonard S. Marks, Corey W. Arnold, Los Angeles, CA
Introduction: MRI-targeted biopsy has improved the accuracy of prostate cancer diagnosis, but regions of interest (ROIs) are often labeled as equivocal category 3 lesions based on PIRADS v2. Here, we demonstrate a machine learning based model for performing "virtual biopsies" with multiparametric MRI (mpMRI) using biopsy Gleason scores as ground truth. This predictive model can be used to assess the probability of the presence of clinically significant cancer (csCaP) at a potential biopsy site using imaging alone.
Methods: Biopsy and mpMRI data from 555 patients from UCLA who underwent MRI-US fusion guided prostate biopsy between 2010 and 2016 were obtained retrospectively. Gleason scores and MR locations for a total of 8,018 biopsy cores were retrieved from these patients. There were 858 cores (10.6%) labeled as containing clinically significant prostate cancer (GS â‰¥ 7). _x000D_
We examined imaging information from regions corresponding to each biopsy location on each MRI parameter (T2, ADC, Ktrans, and calculated high b-value images). Predictive features used to train our machine learning model included texture features (e.g., contrast, homogeneity, etc.) and statistical features (percentiles, mean, and variance). A support vector machine (SVM) classifier with a radial basis function kernel was used, and evaluated using the area under the curve (AUC) statistic.
Results: Data used to train the model included 388 patients using 5-fold cross-validation, with 167 patients held out as a testing set. The classifier achieved an AUC of 0.73 on the test set as well as a mean AUC of 0.70Â±0.02 on the training set. It achieved a negative predictive value of 0.93 and a positive predictive value of 0.23.
Conclusions: In this study, we built a machine learning-based model with moderate (> 0.7) AUC and a high negative predictive value for the presence of clinically significant prostate cancer in biopsy cores using mpMRI alone. This model could be used select targets to improve cancer detection accuracy, and can be further refined through the inclusion of more types of data.
Source of Funding: This work was supported in part by grants from the National Cancer Institute: F30CA210329, R21CA220352, R01CA218547, R01CA195505, and R01CA158627. The authors acknowledge support from the UCLA Computational Integrated Diagnostics program, the UCLA-Caltech Medical Scientist Training Program, and the UCLA Cross-disciplinary Scholars in Science and Technology program. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Cancer Institute or the National Institutes of Health.