Bayesian spatial models for voxel-wise prostate cancer classification using multi-parametric magnetic resonance...

Area of Science:

• Radiology
• Medical Imaging
• Computational Biology

Background:

• Multi-parametric magnetic resonance imaging (mpMRI) is crucial for prostate cancer (PCa) detection.
• Existing computer-aided detection algorithms often overlook spatial correlations and patient heterogeneity in mpMRI data.

Purpose of the Study:

• To develop novel Bayesian approaches for voxel-wise PCa classification.
• To incorporate spatial correlation and between-patient heterogeneity in mpMRI data for improved PCa detection.

Main Methods:

• Proposed three scalable Bayesian methods: Nearest Neighbor Gaussian Process (NNGP), reduced-rank approximation, and conditional autoregressive (CAR) model.
• Modeled spatial correlation and between-patient heterogeneity in mpMRI data.
• Evaluated methods using simulation studies and in vivo data.

Main Results:

• Properly modeling spatial correlation and patient heterogeneity significantly improved PCa classification in simulations.
• All three spatial modeling approaches enhanced classification accuracy on in vivo data.
• The NNGP-based model demonstrated high accuracy and computational efficiency.

Conclusions:

• Spatial modeling in mpMRI data is vital for improving PCa classification.
• The NNGP approach offers a promising, efficient solution for automated PCa detection.