Abstract
Dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) provides a non-invasive tool for evaluating tissue time activity patterns based on the accumulation and metabolism of the contrast agent. Conventional tool for DCE-MRI images analysis is pharmacokinetic model that provides kinetic, physiological parameters for tissues, such as perfusion, capillary permeability, and the volume of extravascular-extracellular space (EES). Although kinetic parameters have shown to be relevant to the response of therapy and the survival rate, inevitable partial volume effect in DCE-MRI images still hinders the quantitative analysis of the kinetic parameters. In this thesis, we develop an unsupervised non-negative blind source separation (nBSS) algorithm to dissect and characterize composite signatures in DCE-MRI images of patients with prostate cancers. We transform the pharmacokinetic model into a latent variable model, to which the nBSS method, named simplex estimation by projection (SIMPLE-Pro) is devised. The SIMPLE-Pro algorithm identifies the tissue time activity curves (up to scaling ambiguity) with theoretical guarantee. The problem of scaling ambiguity is then handled by pharmacokinetic model fitting, which we implemented by using sequential quadratic programming. Some Monte Carlo simulations on synthetic generated prostate DCE-MRI data set and real DCE-MRI experiments of seven patients with prostate cancer were performed to demonstrate high efficiency of SIMPLE-Pro algorithm, and consistency of the extracted information with biopsy test examination. The DCE-MRI analysis framework presented in this thesis, which includes SIMPLE-Pro algorithm and pharmacokinetic model fitting, can dissect complex tissues into regions with differential contrast kinetics at pixel-wise resolution and provide a systems biology tool for defining imaging signatures predictive of phenotypes.