SegWithU: Uncertainty as Perturbation Energy for Single-Forward-Pass Risk-Aware Medical Image Segmentation

Trustworthy Medical Segmentation: Uncertainty-Aware U-Net Evaluation Under Clinical Image Degradation

Medical image segmentation models often report high benchmark accuracy under ideal imaging conditions, yet their failures under clinical degradation can be quiet: sensor noise, patient motion, low- resolution acquisition, and contrast variability may all alter model behavior without producing an obvious warning. We present a reproducible framework for evaluating uncertainty-aware segmentation under con- trolled clinical degradation. Our experiments use a synthetic multimodal brain tumor MRI cohort generated with a biophysical phantom simulator that follows the BraTS protocol. We train U-Net and Attention U-Net baselines for multi-class tumor sub-region segmentation and augment both models with Monte Carlo dropout to estimate per-voxel uncertainty. Across eight clinically motivated corruption types at five severity levels, we measure segmentation accuracy, calibration, failure detection, and selective prediction coverage. On clean data, Attention U-Net achieves a whole-tumor Dice of 0.990; under severe Gaussian noise, its performance falls to 0.089. Predictive uncertainty rises with degradation and tracks segmentation error (Pearson r = 0.53 under severity-3 Gaussian noise), allowing us to flag failures with an AUROC of 0.843. These results argue for uncertainty-aware inference as a practical safety layer in physician-in-the-loop radiology workflows. We release the code, trained models, and evaluation protocol to support direct reproduction.

Uncertainty Quantification in Medical Image Segmentation: A Comprehensive Survey

Uncertainty quantification (UQ) in medical image segmentation is essential for ensuring the reliability and interpretability of deep learning models in clinical decision-making. While convolutional neural networks (CNNs) and transformer-based architectures have achieved remarkable segmentation performance, they often provide deterministic outputs without accounting for uncertainty, which can lead to overconfident predictions in ambiguous cases. This paper presents a comprehensive survey of UQ techniques in medical image segmentation, 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. We examine key methodologies, including Monte Carlo dropout, Bayesian neural networks, variational inference, and ensemble learning, discussing their advantages and limitations in addressing aleatoric and epistemic uncertainties. Additionally, we explore the clinical relevance of UQ by reviewing its applications in brain tumor segmentation, cardiac imaging, lung nodule detection, and other medical domains. The paper also highlights key evaluation metrics, such as calibration errors, uncertainty–error correlation, and visual interpretability, to assess the effectiveness of UQ methods. Finally, we discuss challenges and future research directions, emphasizing the need for scalable, interpretable, and clinically actionable uncertainty quantification strategies to improve trust in AI-assisted medical image analysis.

Semi-supervised Medical Image Segmentation via Perturbation-Aware Mutual Learning and Edge-Aware Uncertainty Loss for Accurate Anatomical Delineation.

Uncertainty-guided mamba network for efficient medical image segmentation with evidential deep learning.

Medical image segmentation requires balancing accuracy, computational efficiency, and uncertainty quantification for potential clinical deployment. Transformer-based architectures achieve superior performance through global context modeling but demand prohibitive computational resources (>20 GFLOPs), while lightweight convolutional networks sacrifice accuracy due to limited receptive fields. We propose an uncertainty-aware efficient segmentation framework synergizing Mamba state-space models with evidential deep learning. Our method employs a 2D-adapted selective state-space mechanism (cross-scan over four directions) to capture long-range dependencies with linear complexity O(L), overcoming transformers' quadratic scaling. The uncertainty-guided attention module (UGAM) leverages Dirichlet-parameterized evidential learning to decompose epistemic and aleatoric uncertainty, adaptively recalibrating features through spatial-channel attention conditioned on prediction confidence. Progressive multi-scale fusion with gradient-based uncertainty supervision enhances boundary delineation and calibration. Experiments on five 2D benchmarks show competitive performance: 82.67% mean Dice on Synapse (1.46% improvement over Swin-UNet, and competitive with state-space peers U-Mamba and Swin-UMamba retrained under the same 2D protocol) with only 7.8M parameters and 4.7 GFLOPs-representing [Formula: see text] parameter reduction and [Formula: see text] efficiency gain. Real-time inference at 37.2 FPS with well-calibrated uncertainty (Expected Calibration Error: 0.046) supports further evaluation of its potential value for time-sensitive clinical workflow analysis rather than direct clinical deployment. A small, single-center reader study using 150 ACDC cases and three radiologists suggested that uncertainty visualization may be associated with improved reader confidence (22.2%) and reduced decision time (18.6%); these exploratory findings require prospective, multi-reader, multi-scanner validation. We explicitly do not claim generalization to 3D volumes, high-resolution pathology, multi-phase CT/MRI, or severe class-imbalance regimes, which are left to future work.

LSR-Diff: A Diffusion Model Synthesizing Level Set Representations for Reliable Segmentation of Medical Images With Ambiguous Edges

Accurate boundary segmentation is critical for high-stakes applications such as disease diagnosis, yet remains challenging due to complex topology, boundary ambiguity, and annotation uncertainty. Diffusion Probabilistic Models (DPMs) generate multiple masks with inherent uncertainty, enhancing boundary delineation compared to deterministic models. However, most existing DPM based segmentation approaches learn discrete binary masks, conflicting with the continuous diffusion process and leading to hard-to-learn degradation during noise addition. Moreover, the prevailing approach of averaging stochastic predictions and applying a fixed threshold disregards structural consistency, often leading to imprecise boundaries, isolated artifacts, and holes. To address these challenges, we propose the <inline-formula> <tex-math notation="LaTeX">$L$ </tex-math></inline-formula>evel <inline-formula> <tex-math notation="LaTeX">$S$ </tex-math></inline-formula>et <inline-formula> <tex-math notation="LaTeX">$R$ </tex-math></inline-formula>epresentation <inline-formula> <tex-math notation="LaTeX">$D$ </tex-math></inline-formula>iffusion model (LSR-Diff), which incorporates a diffusion model with a hybrid mask representation to better capture boundary information, and a novel strategy <inline-formula> <tex-math notation="LaTeX">$E$ </tex-math></inline-formula>nsemble <inline-formula> <tex-math notation="LaTeX">$A$ </tex-math></inline-formula>ggregation via Level <inline-formula> <tex-math notation="LaTeX">$S$ </tex-math></inline-formula>et <inline-formula> <tex-math notation="LaTeX">$E$ </tex-math></inline-formula>volution (EASE) to merge prediction candidates while respecting structural information. The hybrid representation takes advantage of both discrete binary masks and continuous implicit masks, with an intermediate representation to ensure a smooth transition. The EASE module guided by ambiguity estimation and anatomical structure then refines boundary topology, preventing arbitrary mask assembly during the aggregation of stochastic predictions. We conduct extensive experiments across various clinical applications including multiple modalities and tissues, showing that LSR-Diff achieves competitive overall performance and improved edge quality and topology accuracy on the tested tasks.

DynU-Net: Dynamic Uncertainty-Aware Multi-task U-Net for Joint Lesion Segmentation and Classification in Medical Imaging

Segmentation and classification of lesion is essential for computer-aided diagnosis (CAD) in grayscale medical imaging, enabling precise localization and reliable categorization of various pathological conditions. However, conventional multi-task learning (MTL) frameworks often suffer from suboptimal task interactions and require manual loss weight tuning, limiting their effectiveness on challenging modalities such as ultrasound, MRI, and CT. To address these limitations, we propose a Dynamic Uncertainty-aware Network (DynU-Net), a multi-task framework that adaptively balances segmentation and classification through learnable per-task uncertainty parameters. This mechanism eliminates the need for manual hyperparameter tuning while mitigating gradient imbalance during optimization. Extensive experiments on three public datasets, including Brain Stroke CT, Brain Tumor MRI, and BUSI (ultrasound), demonstrate that DynU-Net consistently outperforms both single-task and existing multi-task baselines. In particular, it achieves a Dice score of 88.90% ± 0.33% and a macro-F1 score of 96.73% ± 0.74% on the Brain Stroke CT dataset while maintaining favorable computational efficiency compared to competing architectures. These results highlight the effectiveness of dynamic uncertainty-aware optimization in enabling robust and balanced multi-task learning for joint lesion segmentation and classification on public medical imaging benchmarks.