Abstract:
|
Multiparametric magnetic resonance (MR) imaging represents a powerful tool for developing a non-invasive, user-independent tool for the detection of prostate cancer. Previously, our group showed that a voxel-wise classifier that combined quantitative MR parameters from multiple modalities resulted in better classification of prostate cancer than any single parameter, alone. Anatomically, the prostate can be segmented into multiple zones, with the largest being the central gland and peripheral zone, and a primary limitation of our original model is that we ignored the anatomical structure of the prostate when developing our classifier. In this talk, we discuss approaches to accounting for the anatomical structure of the prostate in a voxel-wise multiparametric classifier for prostate cancer. Two approaches will be discussed: a Bayesian approach, which uses the likelihood to model the anatomical structure of the prostate, and an ensemble learning approach that averages local classifiers from sub-regions of the prostate. We will compare the performance of the two classifiers to each other and to a simple model that does not account for the anatomical structure of the prostate.
|