Functional medical imaging promises powerful tools for thevisualization and elucidation of important disease-causingbiological processes in living tissue. Recent research aims todissect the distribution or expression of multiple biomarkersassociated with disease progression or response, where the signalsoften represent a composite of more than one distinct sourceindependent of spatial resolution. Formulating the task as a blindsource separation or composite signal factorization problem, wereport here a statistically principled method for modeling andreconstruction of mixed functional or molecular patterns. Thecomputational algorithm is based on a latent variable model whoseparameters are estimated using clustered component analysis. Wedemonstrate the principle and performance of the approaches on thebreast cancer data sets acquired by dynamic contrast-enhancedmagnetic resonance imaging.