Gaussian Processes for Enhancing DCE Imaging Quality and Probabilistic Interpretation of Non-Parametric Biomarkers in Soft-Tissue Sarcomas

Motivation: Non-parametric dynamic contrast-enhanced (DCE) MRI lacks a robust framework with which to quantify uncertainty in derived parameters. Goal(s): To develop a framework for statistical treatment of non-parametric measures from DCE-MRI, such as area under the time-intensity curve (AUC) in soft-tissue sarcomas (STS). Approach: Patients with limb STS were scanned using DCE-MRI prior to treatment. A Gaussian Process (radial basis kernel) was applied for voxel-wise signal modelling and prediction; all model parameters were optimized using full Hamiltonian Monte Carlo sampling. Results: GPs effectively model time-varying DCE signals, reducing noise, quantifying non-parametric uncertainties, and enhancing visual quality of dynamic scans. Impact: Gaussian Process modelling of DCE-MRI curves in soft-tissue sarcomas provides uncertainty quantification and reduces image noise, potentially enhancing the characterization of tumour heterogeneity. This approach may offer opportunities for predictive imaging and personalized treatment planning.