Dynamic contrast enhanced magnetic resonance (DCE-MR) imaging is an exciting tool to study the pharmacokinetics of a suspected tumor tissue. Nonetheless, the inevitable partial volume effect in DCE-MR images may seriously hinder the quantitative analysis of the kinetic parameters. In this work, based on the conventional three-tissue compartment model, we propose an unsupervised nonnegative blind source separation (nBSS) algorithm, called time activity curve (TAC) estimation by projection (TACE-Pro), to dissect and characterize the composite signatures in DCE-MR images of patients with prostate cancers. The TACE-Pro algorithm first identifies the TACs (up to a scaling ambiguity) with theoretical support. Then the problem of scaling ambiguity and the estimation of kinetic parameters is handled by pharmacokinetic model fitting. Some Monte Carlo simulations and real DCE-MR image experiments of a patient with prostate cancer were performed to demonstrate the superior efficacy of the proposed TACE-Pro algorithm. Furthermore, the real data experiments revealed the consistency of the extracted information with the biopsy results.