A Pixel-Level Uncertainty-Aware Consistency Learning Framework for Enhancing Feature Compensation In Missing-Modality Glioma Segmentation
Abstract
Purpose
To develop a pixel-level uncertainty-aware consistency (UA-Cons) learning framework to optimize the feature compensation behavior of deep neural networks in scenarios where multi-parametric MRI modalities are incomplete.
Methods
By hypothesizing cross-modal compensation as a confidence-gated process, we developed an Uncertainty-Aware Consistency (UA-Cons) framework, in which missing-modality feature synthesis is governed by a pixel-level uncertainty map derived from Test-Time Augmentation. The mechanisms of 1) feature transfer between multi-modal paths and 2) noise suppression at ambiguous tumor boundaries can thus be optimized through an uncertainty-weighted consistency loss. The UA-Cons framework was validated using 1,251 glioma patients (T1, T1-Ce, T2, FLAIR). Three comparative tiers—baseline, standard consistency, and UA-Cons—were implemented to segment ET, TC, and WT. Performance under severe modality depletion was compared across all tiers, with statistical significance and reliability gains rigorously analyzed using Wilcoxon signed-rank tests.
Results
The UA-Cons framework effectively prevented the model from making wrong guesses by ignoring uncertain signals from the teacher network. Quantitative analysis identified pixel-level uncertainty as the primary reliability metric, with mean uncertainty scores significantly reduced across all tasks: WT (0.4602→0.3037, p<1e-9), TC (0.1448→0.0993, p<1e-9), and ET (0.1476→0.0971, p<1e-9), representing a consistent reduction in predictive noise. Compared to the standard consistency baseline, the UA-Cons mechanism yielded statistically significant Dice coefficient improvements: WT (0.9145→0.9274, p=1.41e-5), TC (0.8922→0.9081, p<1e-6), and ET (0.8542→0.8708, p<1e-6). Furthermore, precision and IoU metrics mirrored these enhancements, particularly in the ET and TC tiers, confirming that the uncertainty-aware gate effectively compensates for feature loss in missing-modality scenarios.
Conclusion
Integrating uncertainty quantification into consistency learning provides a robust solution for incomplete MRI scenarios, simultaneously enhancing segmentation accuracy and quantifiable confidence measures to support reliable clinical decision-making.