Prostate cancer (PCa) is the most common cancer and 2nd leading cause of cancer death among men in the U.S. The use of multiparametric magnetic resonance imaging (mpMRI), a combination of traditional anatomic and newer functional MRI methods, has been shown recently to be helpful for the detection of PCa and identification of clinically significant disease. Although mpMRI is rapidly gaining traction in clinical practice, there are currently no established guidelines for the objective synthesis and evaluation of mpMRI data in widespread use.
The goal of my thesis research is to develop a fully-automated machine learning (ML)-based predictive model that will assess both PCa presence and grade using quantitative mpMRI. The resulting model can in turn be used to aid clinical decision making. Our lab has collected a database of over 50 PCa cases with mpMRI data and co-registered annotated histopathology of radical prostatectomy specimens for each case. The uniqueness of this dataset is that models can be developed on a voxelwise basis rather than requiring the pre-definition of ROIs. While the idea of using computational models for PCa detection is well-established, my proposed model is unique because it will 1) use our data set with a novel ground truth that is more reliable than those of existing models, 2) employ ML methods that will make it more generalizable than standard statistical models, and 3) explicitly aim to determine the extent and grade of PCa.