There has been growing utilization of multi-parametric magnetic resonance imaging (mpMRI) to detect aggressive or high-grade prostate cancer (PCa). Even though close to 20 % of aggressive PCa are missed by mpMRI, this imaging modality is currently being used to guide treatment decisions, such as for surveillance and focal therapy. The goal of this project is to develop a machine learning algorithm leveraging radiomics and genomics to improve the detection of aggressive (Gleason ≥7) PCa. In prior work, we have built radiomics workflows for the characterization of glioblastoma on mpMRI. Additionally, we have developed a targeted next generation sequencing (NGS) assay for genomic characterization of multifocal PCa. We hypothesize that aggressive PCa harbors unique mpMRI features and molecular alterations that can be leveraged to improve detection of aggressive PCa. To test this hypothesis, we propose the following Specific Aims: 1): Develop a machine learning algorithm to detect mpMRI invisible aggressive PCa. Following institutional review board approval, we will identify patients who underwent mpMRI prior to radical prostatectomy (RP) and found to harbor multiple foci of cancer. We will co-register mpMR images with wholemount histopathology to delineate areas of visible and invisible PCa. mpMR images will be pre-processed for spatial registration, corrected for inhomogeneity, segmented, and the respective 3D tumor volumes obtained. 3D graylevel pixel heterogeneity features (fractal texture, Haralick features, and spatial habitats) will be computed from tumor and peritumoral regions. Highly correlated features will be filtered out, and robust features (based on intra-class correlation) retained. Finally, a ‘feature vector’ for gray level heterogeneity will be obtained from each patient’s mpMRI. The cohort will be randomly split into a training, validation and testing sets. Feature selection will be done with the training cohort, followed by model selection in the validation cohort. Finally, the testing set will be used for model assessment. The imaging predictors relevant to MR-visibility will be assessed using feature selection. Based on these selected imaging features, a classification model will be obtained to predict mpMRI visible vs. inviable PCa. The optimal classification model will be selected in the validation set and then evaluated using prediction error, area-under-receiver operative curve (AUROC), and goodness-of-fit statistics in the test set. 2): Create an integrative radiogenomic signature to characterize mpMRI visible PCa. Using patients identified in Aim 1, we will collect the formalin-fixed paraffin-embedded (FFPE) RP specimens harboring multifocal PCa. Targeted multiplexed DNA and RNA PCR-based NGS will be performed to characterize and compare the molecular profiles of visible and invisible PCa foci on mpMRI. A radiogenomic model to predict mpMRI visible lesions combining imaging and genomic alterations will be developed and validated using the approach described in Aim 1.
The successful completion of the proposed project will improve our understanding of the radiogenomic basis of PCa visibility on mpMRI thus facilitating accurate and timely identification of men with aggressive PCa. Additionally, the results may lead to the rapid development of an automated algorithm for interpreting prostate mpMRI.