Bayesian Hierarchical Models for Voxel-Wise Classification of Prostate Cancer Accounting for Spatial Correlation and Between-Patient Heterogeneity in the Multi-Parametric MRI Data

• Activity Number: 470 - Biomarker Evaluation and Winning Student Papers on Medical Devices and Diagnostics
• Type: Contributed
• Date/Time: Wednesday, July 31, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Section on Medical Devices and Diagnostics
• Author(s): Jin Jin* and Joseph Koopmeiners and Lin Zhang and Ethan Leng and Gregory Metzger
• Companies: Division of Biostatistics, University of Minnesota and University of Minnesota and Division of Biostatistics, University of Minnesota and Center for Magnetic Resonance Research, Department of Radiology, University of Minnesota and Center for Magnetic Resonance Research, Department of Radiology, University of Minnesota
• Keywords: Bayesian hierarchical models; Between-patient heterogeneity; Multi-parametric MRI; Multi-image spatial modeling; Nearest Neighbor Gaussian Process; Voxel-wise prostate cancer classi cation

Abstract

Multi-parametric MRI (mpMRI) is an important tool for the diagnosis of prostate cancer. It has some specific features, which have not been fully explored in the literature but could potentially improve cancer detection if leveraged appropriately. We propose a voxel-wise Bayesian classifier that accounts for spatial correlation and between-patient variability in mpMRI. Modeling spatial correlation is challenging due to high dimensionality of the data, and we consider three computationally efficient approaches through Nearest Neighbor Gaussian Process (NNGP), reduced-rank approximation, and a conditional autoregressive (CAR) model. The between-patient variability is accounted for by a subject-specific random intercept on mpMRI parameters. Simulation results demonstrate that properly modeling the aforementioned features of mpMRI improves classification. Application to in vivo data illustrate that classification is improved by spatial modeling using NNGP and reduced-rank approximation but not the CAR model, while modeling between-patient variability does not further improve our classifier. The model using NNGP is recommended, with robust performance and high computational efficiency.