Background: Breast cancer heterogeneity is thought to be associated with adverse outcomes. Dynamic molecular imaging modalities, including PET, permit 4-D sampling of tumor biologic properties and can therefore capture functional heterogeneity revealed by the temporal dimension of the dynamic tracer uptake. With the goal of improved non-invasive characterization of in vivo tumor biology, we have developed and tested a novel radiomic biomarker that characterizes 4-D functional tumor heterogeneity (FTH). We hypothesize subclonal populations are spatially contiguous regions of shared physiologic behavior that drive breast cancer heterogeneity and can be quantified. We describe an initial application of this approach to FDG PET imaging of breast cancer. Methods: We retrospectively analyzed data from a study of 50 patients with locally advanced breast cancer. Patients underwent 60-minute dynamic FDG PET over the chest prior to neoadjuvant chemotherapy and breast surgery and were followed for disease recurrence (DFS). A 3-D region bounding each tumor was identified by a radiologist and a novel Markov Random Field based 4-D segmentation paradigm segmented each tumor into three spatially constrained sub-regions with distinct time activity curve dynamics. From each tumor, an FTH imaging signature was extracted characterizing cluster compactness and separation. FTH imaging signatures were z-score normalized across all patients. Time-to-event analysis was used to assess the prognostic value of the FTH imaging signatures in predicting DFS. Discriminatory capacity compared to a baseline model of established prognostic factors (age, hormone receptor status, baseline tumor size) and standard PET uptake and kinetics markers (SUV, K₁, and Ki) shown to be predictive of DFS (Dunnwald, J Clin Oncol 26:4449, 2008) was evaluated using the c-statistic and the log-likelihood statistical test. Results: 17 of 50 women (34%) had recurrence events. Adding FTH imaging signatures to the baseline model of age, baseline tumor size, and hormone receptor status improved a cross validated c-statistic from 0.51 to 0.74 (p < 0.05), and demonstrated higher discriminatory capacity over a model of age, tumor size, hormone receptor status, and standard PET measures (c-statistic = 0.59). Conclusions: Imaging biomarkers of 4-D metabolic tumor heterogeneity may add prognostic value in predicting recurrence-free survival in breast cancer and merit further study.