Confidence-weighted ensemble strategies based on task-informed data-augmentation to improve segmentation robustness: application to kidney MRI

Artikel i vetenskaplig tidskrift, 2026

Deep learning–based segmentation models have achieved remarkable performance in medical image analysis, yet their deployment in clinical practice is hindered by limited robustness and unreliable uncertainty estimates under distribution shifts. In renal magnetic resonance imaging (MRI), variability due to geometric transformations, intensity inhomogeneities, and resolution loss can substantially degrade model reliability. In this study, we propose an uncertainty-aware ensemble framework that combines task-informed data augmentation with confidence-weighted prediction aggregation to improve MRI renal segmentation robustness and reliability. Multiple U-Net models equipped with Monte Carlo Dropout are trained using distinct, realistic MRI-specific augmentation strategies, each targeting a specific type of natural variation (affine transformations, bias field artifacts, and blurring). At inference time, predictions are aggregated using ensemble strategies that explicitly leverage predictive uncertainty to weight or select the most reliable outputs. Model performance is evaluated on a renal MRI dataset comprising healthy controls and chronic kidney disease patients, under controlled perturbations reflecting realistic clinical distortions. Robustness is assessed through a unified framework that jointly considers segmentation accuracy and multiple uncertainty-based metrics. Results demonstrate that confidence-weighted ensemble strategies consistently outperform single-model and conventional soft-voting baselines, achieving higher robustness across all perturbation types. The proposed approach enhances error detection and supports more informed human–AI collaboration by identifying unreliable predictions. Overall, these findings indicate that selectively exploiting uncertainty within ensembles of task-specialized models is a more effective strategy than relying on comprehensive augmentation within a single model, offering a robust and clinically relevant solution for renal MRI segmentation.