HUR-MACL: Uncertainty-Guided Collaborative Segmentation

• The paper presents a novel framework that dynamically identifies high-uncertainty regions and deploys complementary architectures to boost segmentation accuracy.
• It combines global context from Vision Mamba with local boundary refinement from Deformable CNN to precisely segment challenging anatomical structures.
• Uncertainty-guided distillation and adaptive loss modulation enable efficient training protocols that outperform conventional single-network approaches on complex datasets.

High Uncertainty Region-Guided Multi-Architecture Collaborative Learning (HUR-MACL) is a framework for medical image segmentation that employs region-specific uncertainty estimation to orchestrate collaboration among heterogeneous deep learning architectures. Its design focuses computational effort on the most ambiguous or difficult anatomical regions, combining the strengths of multiple segmentation networks while minimizing functional overlap and redundant computation. The following article synthesizes key implementations and results from HUR-MACL models across recent literature, including architectural innovations, region mining techniques, collaborative learning strategies, and validated experimental outcomes (Lu et al., 15 Dec 2025, Zheng et al., 2021, Liu et al., 8 Jan 2026).

1. Definition and Problem Setting

HUR-MACL targets the segmentation of complex medical structures where irregular shapes, small size, or low contrast yield high prediction uncertainty. Conventional single-network or hybrid approaches, which treat the entire image uniformly, often underperform in these “hard” regions and fail to exploit the unique representational strengths of different model families. HUR-MACL frameworks address this by:

• Dynamically identifying regions of high epistemic or aleatoric uncertainty.
• Selectively deploying multiple, complementary architectures (e.g., Vision Mamba, Deformable CNN, ViT-based teachers, CNN students) to those regions.
• Enforcing feature or output-level distillation between architectures in the mined uncertain areas.
• Modulating loss functions by uncertainty so that learning and supervision are focused where model confidence is lowest.

Such region-guided collaboration yields demonstrable improvements in segmentation accuracy and boundary adherence, particularly for challenging organ-at-risk (OAR) structures in medical datasets (Liu et al., 8 Jan 2026).

2. High-Uncertainty Region Mining

Identification of “hard” regions is central to HUR-MACL architectures. The process typically involves:

• Computing per-pixel class probability maps from a backbone encoder–decoder (e.g., U-Net).
• Calculating normalized Shannon entropy maps:

$$U_i(h,w) = -\frac{1}{\log |C|} \sum_{c=1}^{C} y_i(h,w,c) \log y_i(h,w,c)$$

for decoder level $i$, where $y_i$ are softmax outputs and $C$ is class count (Liu et al., 8 Jan 2026).

• Applying a threshold $T$ (typically $10^{-3}$) to generate a binary mask $M_i(h, w) = \mathbb{1}[U_i(h,w) \ge T]$, which selects “hard” pixels.
• Passing only masked high-uncertainty regions on to downstream collaborative architectures.

Other variants utilize Monte-Carlo dropout-based predictive entropy (Zheng et al., 2021) or multi-teacher epistemic and mutual information metrics (Lu et al., 15 Dec 2025) to quantify uncertainty. These methods consistently demarcate error-prone or anatomically complex areas, allowing focused allocation of representational power.

3. Multi-Architecture Collaborative Segmentation

In HUR-MACL, the mined high-uncertainty regions are subject to processing by two or more distinct architectures:

• Vision Mamba (ViM) Encoder: Operates at patch level, applying state-space models and token-wise gating to incorporate global context. Masked region features are embedded, gated, and fused via LayerNorm and SiLU nonlinearity:

$$S_i' = \mathrm{LayerNorm}\left(\mathrm{SiLU}(z) \odot y_{\text{forw}} + \mathrm{SiLU}(z) \odot y_{\text{back}}\right)$$

leading to refined segmentation in “hard” regions (Liu et al., 8 Jan 2026).

• Deformable CNN (DCNN) Encoder: Processes the same masked features with convolutional layers that learn offset fields, adapting receptive fields to object shapes:

$$(f *_{\text{def}} k)(p) = \sum_{q \in \mathcal{R}} k(q) f\left(p + q + \Delta x(p)\right)$$

yielding segmentations that excel in capturing local boundary complexity (Liu et al., 8 Jan 2026).

• Dual-Teacher Strategies: In semi-supervised variants, the generalized teacher (frozen ViT-based foundation model) distills broad priors while the specialized teacher (EMA of the student) adapts to fine-grained domain specifics (Lu et al., 15 Dec 2025). Collaborative learning is achieved via dual-path knowledge distillation and uncertainty-aware pseudo-label fusion.

4. Uncertainty-Guided Distillation and Loss Modulation

Cross-architecture supervision in HUR-MACL is locally modulated by model confidence via pixel-wise bidirectional loss functions.

• Feature Distillation in Reliable Regions: For each high-uncertainty region, the outputs of Vision Mamba and DCNN are compared via cross-entropy:

$$\mathrm{CE}^M(h,w) = -\sum_c M_{\mathrm{GT}}(h,w, c) \log P^M(h,w, c)$$

Direction flags $m(h,w)$ guide teaching: each model supervises the other only at pixels where it is more accurate.

• Bidirectional KL Minimization:

$$L_P^M = \frac{1}{\sum(1-m)}\sum_{h,w}(1-m(h,w))\, \mathrm{KL}\left(P^M(h,w) \| P^D(h,w)\right)$$

$$L_P^D = \frac{1}{\sum m}\sum_{h,w}m(h,w)\, \mathrm{KL}\left(P^M(h,w) \| P^D(h,w)\right)$$

with total distillation loss $loss_3 = L_P^M + L_P^D$ (Liu et al., 8 Jan 2026).

• Soft Uncertainty Weighting (in pseudo-label supervised learning): Loss terms are weighted by $w(u) = \exp(-u)$, steering optimization away from ambiguous voxels and toward the most confidently labeled regions (Lu et al., 15 Dec 2025, Zheng et al., 2021).

The training objective aggregates baseline segmentation loss, hard-region collaborative loss, and distillation loss:

$$\mathcal{L} = \mathcal{L}_{\text{seg}}\left(P^U,\, \mathrm{GT}\right) + \alpha\, loss_2 + \beta\, loss_3$$

with $\alpha=1.0$, $\beta=0.5$, and balanced Dice/cross-entropy weighting (Liu et al., 8 Jan 2026).

5. Training Protocols and Hyperparameterization

Training in HUR-MACL is structured around multi-stage optimization and uncertainty-adaptive schedules:

• Backbone and Branch Configurations:
  • U-Net backbone for full-image segmentation.
  • Vision Mamba and DCNN branches applied to mined hard regions.
• Two-Stage Training (Lu et al., 15 Dec 2025):
  • Stage 1: Pretraining with labeled data; dual-path model distillation.
  • Stage 2: Semi-supervised fine-tuning; pseudo-label loss modulated by uncertainty, only visual distillation retained from the foundation teacher.
• Hyperparameters:
  • Entropy threshold $T = 10^{-3}$ for region masking (Liu et al., 8 Jan 2026).
  • $\lambda_{\mathrm{Dice}}$, $\lambda_{\mathrm{CE}}$, $\alpha$, $\beta$ empirically set for balanced convergence.
  • EMA momentum $\mu=0.99$ in dual-teacher variants.
  • Soft uncertainty weighting parameter $\alpha=1$ in most instantiations (Lu et al., 15 Dec 2025).
• Optimization Details:
  • SGD with Nesterov momentum (0.99) and learning rate decay (Liu et al., 8 Jan 2026).
  • Monte-Carlo dropout (rate $p=0.5$, $T=10$ samples per image) for uncertainty estimation in some versions (Zheng et al., 2021).

6. Quantitative Results and Benchmarks

Performance metrics substantiate HUR-MACL’s efficacy, especially on “hard” anatomy and with low annotation rates.

Method	PDDCA Dice ↑	PDDCA ASSD ↓	StructSeg Dice ↑	StructSeg ASSD ↓	Inhouse Dice ↑	Inhouse ASSD ↓
U-Net	75.95 %	1.32 mm	72.92 %	1.38 mm	72.34 %	1.81 mm
nnU-Net	76.64 %	1.56 mm	73.60 %	1.50 mm	72.42 %	1.62 mm
FocusNet	77.23 %	1.78 mm	74.53 %	1.42 mm	70.63 %	1.97 mm
HUR-MACL	81.84 %	1.16 mm	78.32 %	1.08 mm	74.11 %	1.72 mm

Additional results:

• Optic chiasm Dice rises from $\sim$56% (FocusNet) to 66% on PDDCA.
• Semi-supervised HUR-MACL with 5–10% labels can outperform fully supervised and zero-shot baselines on brain MR, pancreas CT, and thoracic CTA datasets (Lu et al., 15 Dec 2025).
• Monte-Carlo dropout uncertainty weighting yields $\sim$2% DSC gain in supervised stages and $\sim$1.4% in unsupervised agreement (Zheng et al., 2021).
• Ablation shows feature distillation in mined regions further boosts overall performance by up to 2.7% DSC on small anatomical structures (Liu et al., 8 Jan 2026).

7. Methodological Insights and Practical Guidelines

Critical components and insights include:

• Region-guided mining sharply focuses computation where models are least confident, maximizing synergistic gains from multi-architecture ensembles.
• Vision Mamba encoders address global shape and context, while Deformable CNNs capture local boundary irregularities—fusion of these produces superior contour refinement.
• Distillation losses restricted to “light” pixels ensure mutual architectural improvement and prevent collapse to similar, potentially suboptimal, representations.
• Threshold settings for hard-region selection are critical; over-expansion introduces redundancy and degrades accuracy (Liu et al., 8 Jan 2026).
• In dual-teacher regimes, generalized teachers control for robust priors; specialized teachers adapt to clinical heterogeneity; uncertainty gating balances their supervision (Lu et al., 15 Dec 2025).
• Joint backpropagation of region-specific and global objectives accelerates and stabilizes training convergence (Zheng et al., 2021).

A plausible implication is that such uncertainty-guided multi-architecture collaboration generalizes to other domains requiring detailed, context-aware segmentation under annotation constraints, given robust region mining and distillation techniques.

• Harmonizing Generalization and Specialization: Uncertainty-Informed Collaborative Learning for Semi-supervised Medical Image Segmentation (Lu et al., 15 Dec 2025)
• Uncertainty-Aware Deep Co-training for Semi-supervised Medical Image Segmentation (Zheng et al., 2021)
• HUR-MACL: High-Uncertainty Region-Guided Multi-Architecture Collaborative Learning for Head and Neck Multi-Organ Segmentation (Liu et al., 8 Jan 2026)