Spectral clustering applied for dynamic contrast-enhanced MR analysis of time-intensity curves

Dynamic contrast-enhanced (DCE)-magnetic resonance imaging (MRI) represents an emerging method for the prediction of biomarker responses in cancer. However, DCE images remain difficult to analyze and interpret. Although pharmacokinetic approaches, which involve multi-step processes, can provide a general framework for the interpretation of these data, they are still too complex for robust and accurate implementation. Therefore, statistical data analysis techniques were recently suggested as another valid interpretation strategy for DCE-MRI. In this context, we propose a spectral clustering approach for the analysis of DCE-MRI time-intensity signals. This graph theory-based method allows for the grouping of signals after spatial transformation. Subsequently, these data clusters can be labeled following comparison to arterial signals. Here, we have performed experiments with simulated (i.e., generated via pharmacokinetic modeling) and clinical (i.e., obtained from patients scanned during prostate cancer diagnosis) data sets in order to demonstrate the feasibility and applicability of this kind of unsupervised and non-parametric approach.