UCBound-Net is introduced, a continual segmentation framework that exploits Monte Carlo (MC) Dropout uncertainty as a spatial proxy for forgetting risk and outperforms baseline methods without requiring task-boundary supervision.
Abstract
Continual learning in clinical imaging faces a dual challenge: a model must assimilate knowledge from new anatomical domains while retaining representations learned from prior tasks, a problem known as catastrophic forgetting. Existing mitigation strategies, including regularization and knowledge distillation, treat all spatial regions equally, ignoring the fact that prediction uncertainty is strongly correlated with the propensity for forgetting. We introduce UCBound-Net, a continual segmentation framework that exploits Monte Carlo (MC) Dropout uncertainty as a spatial proxy for forgetting risk. Our method contributes three synergistic components: (i) uncertainty-weighted boundary distillation, which amplifies the knowledge transfer signal at high-entropy regions of the frozen teacher; (ii) uncertainty-calibration regularization, which explicitly penalizes overconfident erroneous predictions; and (iii) uncertainty-guided exemplar selection, a memory buffer that preferentially stores samples whose boundary regions exhibit the highest predictive entropy. Evaluated on a sequential domain-incremental benchmark comprising breast ultrasound (BUSI, Task 1) followed by thyroid ultrasound (TN3K, Task 2), UCBound-Net reduces forgetting relative to naive fine-tuning, achieving a backward transfer (BWT) of -0.098 compared with -0.173, while obtaining an average Dice Similarity Coefficient (DSC) of 0.755 across both tasks. The proposed framework outperforms baseline methods without requiring task-boundary supervision. An ablation study further demonstrates that each component contributes independently to forgetting mitigation, providing a practical pathway toward uncertainty-aware continual learning for clinical image segmentation.
This work proposes an uncertainty-aware efficient segmentation framework synergizing Mamba state-space models with evidential deep learning, employing a 2D-adapted selective state-space mechanism to capture long-range dependencies with linear complexity O(L), overcoming transformers' quadratic scaling.
A comprehensive survey of UQ techniques in medical image segmentation is presented, categorizing existing approaches into Bayesian methods, deep ensembles, deterministic methods, test-time data augmentation, and hybrid models, while treating foundation-model-based UQ as a separate cross-cutting category.
Seyed Sina Ziaee, K. Ovens· Journal of Imaging· 0 citations
Encoder-decoder segmentation networks use skip connections to recover spatial detail, but direct feature transfer can also propagate irrelevant high-frequency responses into the decoder. This problem is pronounced at weak or irregular tumor margins, where contour evidence is useful but should not modify semantically reliable features indiscriminately. We propose Bridge-Net, a boundary-ambiguity guided residual injection framework for two-dimensional brain tumor MRI segmentation. Bridge-Net formulates boundary enhancement as a conditional feature-correction problem. A structural boundary prior is coupled multiplicatively with an ambiguity map obtained from an auxiliary foreground logit. The ambiguity cue is deterministic and probability-based, rather than a Bayesian or calibrated uncertainty estimate. The overlap between the two cues identifies locations that are both boundary-like and prediction-ambiguous. A level-specific gated residual branch then injects the resulting cue into each skip feature, while a zero-initialized bounded scaling factor preserves an identity-like main pathway at the start of optimization. Experiments were conducted on TCGA-LGG and BRISC2025 under dataset-specific protocols, including patient-level partitioning for TCGA-LGG and the official image-level split for BRISC2025. Bridge-Net achieved Dice/IoU/HD95 values of 84.16%/73.65%/15.38 on TCGA-LGG and 87.26%/77.63%/8.68 on BRISC2025. Patient-level paired analysis on TCGA-LGG and image-level paired analysis on BRISC2025 further supported the improvements over UCTransNet. Ablation results support the complementary roles of structural boundary and probability-ambiguity cues, and a same-seed repeated-run comparison across reproduced models supports the stability of the observed performance trend under the tested setting. Relative to UCTransNet, the proposed mechanism increases the parameter count from 7.982 M to 7.988 M and FLOPs from 24.083 G to 24.196 G at 256×256 resolution. These results indicate that ambiguity-filtered boundary residual fusion introduces only a small increase in parameter count and FLOPs for boundary-ambiguous MRI segmentation, although broader patient-level and volumetric validation remains necessary.
Haoran Gu, Shuo Guo, Yuhan Ying et al.· Mathematical and Computation...· 0 citations
A Dynamic Uncertainty-aware Network (DynU-Net) is proposed, a multi-task framework that adaptively balances segmentation and classification through learnable per-task uncertainty parameters that consistently outperforms both single-task and existing multi-task baselines.
Ngoc Ly Tran, Thi Thu Thuy Nguyen, Ba-Hung Ngo et al.· Journal of Computational Des...· 0 citations
These findings expose a fundamental mismatch between the theoretical promise of aleatoric uncertainty and its practical behavior, and suggest that practitioners should not rely on entropy-based uncertainty as a proxy for clinical ambiguity in safety-critical applications.
Simon Baur, Arne Schernich, Ekin Böke et al.· 1 citation
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
