Uncertainty Estimation in Deep Learning MRI Reconstruction with Focus on Pathologies

Title: Uncertainty Estimation in Deep Learning MRI Reconstruction with Focus on Pathologies Authors: Vanya Saksena, Florian Knoll, Bernhard Kainz Abstract: Deep learning (DL) enables high-quality accelerated MRI reconstruction, yet clinical adoption remains limited due to undetected reconstruction errors such as hallucinated features, subtle image alterations, and missed pathologies. While uncertainty estimation (UE) aims to identify unreliable regions, most evaluations focus on global image quality rather than diagnostically relevant structures. We systematically investigate whether uncertainty maps correlate with reconstruction errors at the pathology level using the fastMRI+ brain dataset. Uncertainty maps are generated using two approaches: pathology-aware model ensembling and Monte Carlo sampling via input noise perturbation. To simulate realistic clinical conditions, models are trained on limited data to induce suboptimal reconstructions. Across multiple pathology types and acceleration factors, model ensembling demonstrates strong linear correlation between uncertainty and reconstruction error, whereas noise perturbation shows limited sensitivity.