Abstract:
|
Multi-parametric magnetic resonance imaging (mpMRI) has recently received increased attention as a non-invasive method for the detection of prostate cancer. Due to the subjective nature of standard diagnostics, its full potential in patient management hasn't been fully realized. Previously, a quantitative, user-independent, multi-parametric classification model was developed. However, it ignored the fact that various prostate regions are associated with both mpMRI parameter values and voxel-wise cancer risk. In this paper, we propose a novel classifier for prostate cancer detection under a Bayesian framework, which aims to improve the classification accuracy by accounting for the anatomic structure of the prostate. We also combine the classifier with a spatial smoother to account for residual spatial correlation in the data. Results show that our proposed model achieved significant improvements in prostate cancer detection compared to a baseline model which did not consider the structure of the prostate. The closed-form solution for the posterior cancer probability also helps incorporate further model structures with a minimal loss in computational efficiency.
|