Uncertainty Estimation Through Test-Time Augmentation in Deep Learning for Arterial Input-Function Estimation

Abstract

Uncertainty estimation is an important component in the safe deployment of deep learning (DL) models, particularly in high-stakes domain such as medical imaging. While DL models often achieve impressive predictive performance, they typically provide no indication of how reliable their outputs are. In this thesis, we investigate the use of test-time augmentation (TTA) to estimate predictive uncertainty in DL models for arterial input function (AIF) prediction from 4D dynamic PET data. We hypothesize that the variability in model predictions under input perturbations can be used to flag unreliable outputs. Our proposed method leverages domain-specific TTA strategies that reflect realistic sources of variability in PET imaging, and requires no model retraining or architectural changes, making it practical for clinical workflows. We show that incorporating domain knowledge into augmentation strategies improves the informativeness of the resulting uncertainty estimates. By evaluating our framework on small-animal PET data, we find that TTA-based uncertainty can effectively flag a subset of inaccurate predictions. The method achieves high precision in identifying high-error cases, but suffers from low recall - successfully detecting some unreliable outputs while missing others. These findings suggest that TTA offers a simple and interpretable approach to uncertainty estimation, though further refinement is needed to ensure robust error detection across all samples.