UnCoL Framework: Dual-Teacher Segmentation

• UnCoL Framework is an uncertainty-informed dual-teacher semi-supervised approach that integrates generalized and specialized learning for precise medical image segmentation.
• It employs dual-path knowledge distillation and pixel-level uncertainty gating to fuse prompt-conditioned and EMA teacher guidance for improved pseudo-labeling.
• Empirical results on 2D and 3D datasets demonstrate that UnCoL achieves near fully supervised performance with significantly fewer annotations.

The Uncertainty-informed Collaborative Learning (UnCoL) framework is a dual-teacher semi-supervised approach designed to harmonize generalization and specialization for medical image segmentation under limited annotation. UnCoL distills knowledge from both a frozen, prompt-conditioned foundation model and a task-adaptive, exponentially averaged teacher to guide a student model. Its training pipeline leverages explicit uncertainty modeling to regulate pseudo-label supervision, thereby suppressing unreliable guidance and stabilizing learning in ambiguous regions. This architecture yields consistent improvements over both classic and modern semi-supervised segmentation methods, approaching fully supervised performance with reduced annotation requirements (Lu et al., 15 Dec 2025).

1. Dual-Teacher Architecture and Training Workflow

UnCoL comprises three major model components:

• Generalized Teacher ($f_\xi$): A prompt-conditioned, frozen segmentation foundation model (e.g., MedSAM or SAM-Med3D) that provides large-scale semantic and visual priors. Parameters $\xi$ remain fixed throughout training.
• Specialized Teacher ($f_{\theta_S}$): An exponential moving average (EMA) clone of the student model, parameters updated by

$$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$

adapting continuously to domain- and task-specific idiosyncrasies.

• Student Model ($f_\theta$): A lightweight, prompt-free segmentation network (typically SimpleViT encoder plus U-Net or V-Net decoder) trained to absorb both broad generalization priors and dataset-specific structure.

The UnCoL training process is divided into two stages:

• Pretraining on labeled data: Student is trained with full supervision ($\mathcal{L}_{\rm sup}$) and dual-path knowledge distillation (DPKD) from the Generalized Teacher.
• Semi-supervised Fine-tuning on labeled ($\mathcal{D}_L$) and unlabeled ($\mathcal{D}_U$) data: Continues $\mathcal{L}_{\rm sup}$, maintains visual distillation, and introduces Uncertainty-Aware Pseudo-Labeling (UAPL) that adaptively integrates pseudo-labels from either teacher depending on estimated confidence.

During fine-tuning, both teachers output class-probability maps $p^G$, $\xi$0 and per-pixel entropy-based uncertainty $\xi$1, $\xi$2. At each spatial position $\xi$3, a mask $\xi$4 identifies teacher $\xi$5 as confident if the uncertainty is below schedule $\xi$6. The pseudo-probabilities $\xi$7 are computed by uncertainty-weighted fusion; pseudo-labels $\xi$8 are then used for student supervision over reliable spatial regions.

2. Dual-Path Knowledge Distillation

To transfer rich generalization capacity from the foundation model, UnCoL implements DPKD with two complementary losses:

• Visual Distillation aligns intermediate ViT representations,

$\xi$9

where $f_{\theta_S}$0, $f_{\theta_S}$1 are teacher/student features, and $f_{\theta_S}$2 projects student features to the teacher embedding space.

• Semantic Distillation aligns final fusion outputs:

$f_{\theta_S}$3

with $f_{\theta_S}$4 a learned linear map, $f_{\theta_S}$5 the final student encoder output, and $f_{\theta_S}$6 the prompt-fused teacher output.

The total distillation loss is

$f_{\theta_S}$7

Visual distillation is sustained throughout training, while semantic distillation is disabled during semi-supervised fine-tuning to avoid unreliable prompt signals.

3. Uncertainty-Aware Pseudo-Label Learning

UnCoL's Uncertainty-Aware Pseudo-Labeling mechanism is regulated at the pixel level by per-teacher confidence:

• Uncertainty Estimation: Teacher confidence at pixel $f_{\theta_S}$8 is assessed as Shannon entropy,

$f_{\theta_S}$9

for $$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$0.

• Threshold Schedule: Ramp-up threshold $$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$1, where $$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$2, promotes conservative supervision early and gradually admits more ambiguous pixels.
• Fusion: Where both teachers are confident, predictions are blended via exponential-entropy weighting:

$$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$3

If only one teacher is confident, only its $$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$4 is used; otherwise, supervision is excluded for that pixel.

For pseudo-label loss, valid regions $$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$5 are selected. The student is supervised using hybrid cross-entropy and Dice:

$$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$6

4. Training Objective and Hyperparameterization

Loss composition is adjusted by phase:

• Pretraining (labeled only):

$$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$7

where $$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$8, $$\theta_S \leftarrow \mu\,\theta_S + (1-\mu)\,\theta,\quad \mu=0.99,$$9.

• Semi-supervised fine-tuning:

$f_\theta$0

with $f_\theta$1, $f_\theta$2, $f_\theta$3.

Optimization uses SGD (lr = 0.01), weight decay ($f_\theta$4), 15,000 iterations per stage, EMA momentum $f_\theta$5, and batch sizes 4 ($f_\theta$6 labeled, $f_\theta$7 unlabeled). Spatial copy–paste augmentation further enhances sample diversity. Inference requires a single forward pass through the prompt-free student model.

5. Experimental Results and Empirical Performance

UnCoL achieves superior segmentation accuracy compared to zero-shot foundation models, classical and contemporary semi-supervised learning (SSL) baselines. On 2D OASIS with 5% labels, UnCoL reaches $f_\theta$8 Dice (vs. $f_\theta$9 zero-shot MedSAM, $\mathcal{L}_{\rm sup}$0 full-sup UNet). For 3D Pancreas-CT with 10% labels, UnCoL yields $\mathcal{L}_{\rm sup}$1 Dice (vs. $\mathcal{L}_{\rm sup}$2 other SSL and $\mathcal{L}_{\rm sup}$3 MedSAM-3D zero-shot). On 3D ImageTBAD with 20% labels, UnCoL attains Dice $\mathcal{L}_{\rm sup}$4, correcting errors not rectified by either individual teacher.

Uncertainty measures are well-calibrated (AUROC $\mathcal{L}_{\rm sup}$5, ECE $\mathcal{L}_{\rm sup}$6), reliably discriminating correct from incorrect regions. Ablation confirms that neither frozen nor EMA teacher alone suffices: only their uncertainty-gated combination yields top performance in both accuracy and boundary delineation (metrics 95HD, ASD).

6. Significance and Representational Impact

UnCoL formally harmonizes generalization (via frozen foundation knowledge distillation) and specialization (via EMA adaptation) while stabilizing pseudo-label learning through pixel-wise uncertainty gating. This approach addresses domain shift, data scarcity, and inter-task ambiguity typical in medical image segmentation. The explicit uncertainty mechanism mitigates confirmation bias and propagation of erroneous pseudo-labels in unlabeled regions.

A plausible implication is that the UnCoL framework's dual-teacher and uncertainty-gated design pattern could generalize to other domains where tension between broad transfer and local adaptation is critical. Its modular structure permits integration with modern segmentation backbones and foundation models.

7. Summary Table

Component	Description	Role
Generalized Teacher	Frozen, prompt-based foundation model	Semantic prior
Specialized Teacher	EMA of student	Domain adaption
Student Model	SimpleViT + U/V-Net	Target learner
Pseudo-label Strategy	Uncertainty-weighted, per-pixel	Gated learning
Distillation Pathways	Visual and semantic	Representation

UnCoL's dual-teacher, uncertainty-aware formulation sets a new methodological baseline for semi-supervised segmentation, especially in settings characterized by limited labeled data and diverse annotation regimes (Lu et al., 15 Dec 2025).