UP2D: Uncertainty-aware Progressive Pseudo-label Denoising for Source-Free Domain Adaptive Medical Image Segmentation

Quang-Khai Bui-Tran, Thanh-Huy Nguyen, Ho Dung Manh, Thinh B. Lam, Vi Vu, Hoang-Thien Nguyen, Phat K. Huynh, Ulaş BağcıarXiv 2025 · DOI · arXiv

TL;DR

UP2D is a new framework for medical image segmentation that adapts to new datasets without ever needing to see the original training data—a critical requirement for patient privacy. By combining a student-teacher architecture with a clever "uncertainty-guided" update strategy and refined label denoising, it achieves state-of-the-art accuracy on retinal fundus images, even when facing significant domain shifts.

What Problem Does This Paper Address?

Deep learning models for medical imaging often fail when deployed at a new hospital because of "domain shift"—differences in camera hardware, lighting, or patient demographics. Traditionally, we fix this using Unsupervised Domain Adaptation (UDA), which requires access to both the original (source) and new (target) data.

However, strict privacy laws (like HIPAA or GDPR) often make sharing source medical data impossible. This leads us to Source-Free Domain Adaptation (SFDA), where we only have a pre-trained model and unlabeled target images. The challenge here is that the model must learn from its own "pseudo-labels," which are often noisy and incorrect. If the model trains on its own mistakes, it enters a death spiral of "confirmation bias," eventually degrading in performance.

Methods

The authors introduce UP2D (Uncertainty-aware Progressive Pseudo-label Denoising), a single-stage student-teacher framework. The teacher generates pseudo-labels, and the student learns from them. To prevent the "death spiral," they introduce three novel components:

  1. Refined Prototype Filtering (RPF): Instead of trusting every pixel in a pseudo-label, the model calculates "class prototypes" (average feature representations). It then filters out pixels that are too far from their class prototype or have high uncertainty (measured via Monte Carlo Dropout). This is especially helpful for small, difficult structures like the optic cup.
  2. Uncertainty-Guided EMA (UG-EMA): Usually, a teacher model is updated by a simple moving average of the student's weights. UP2D is pickier: it only updates the teacher when the student achieves a new "record low" in uncertainty. It uses an Inverted-Gaussian weighting map to prioritize learning at the boundaries of objects, where errors are most likely to occur.
  3. Quantile-based Entropy Minimization: Standard entropy minimization forces the model to be confident everywhere. UP2D uses a quantile filter to ignore pixels that are already very certain (high confidence) or clearly noise. This forces the model to focus its learning energy on the "ambiguous" regions that actually need improvement.

Datasets

The researchers evaluated UP2D on three major retinal fundus image benchmarks for optic disc and optic cup segmentation:

  • REFUGE: Used as the source domain (320 images).
  • Drishti-GS: A target domain dataset (50 training / 51 testing images).
  • RIM-ONE-r3: A second target domain dataset (90 training / 60 testing images).
  • REFUGE Validation: Used to test "Open-Domain" performance (unseen data during adaptation).

Key Insights and Ideas

  • Boundaries Matter Most: By using an Inverted-Gaussian map, the authors realized that the center of an object is usually easy to segment. The real "knowledge" is at the edges. Prioritizing these regions during the teacher update significantly boosts boundary precision.
  • Selective Memory Beats Constant Learning: The UG-EMA strategy acts as a quality control gate. By only updating the teacher when the student provides "reliable feedback" (low uncertainty), the system avoids incorporating the student's temporary errors into the long-term "teacher" knowledge.
  • Denoising Small Objects: In fundus images, the "optic cup" is much smaller than the "optic disc." Standard denoising often wipes out the cup entirely. The RPF module specifically masks out dominant background signals to preserve features for these underrepresented classes.

Key Results

UP2D outperformed several high-profile UDA and SFDA methods (like BEAL, TENT, and CPR) across multiple metrics:

  • Drishti-GS: Achieved a 96.61% Dice score for the optic disc and 86.61% for the optic cup, narrowing the gap with "Target Only" models (models trained on labeled target data).
  • RIM-ONE-r3: Reached a 95.17% Dice score for the disc, significantly outperforming prior SFDA methods like DPL (85.98%).
  • Open-Compound Domain: In scenarios where the model encountered a mix of seen and unseen target domains, UP2D maintained superior generalization, outperforming the next best method by over 2% in Dice score.
  • Qualitative Improvement: Visual results show much cleaner boundaries and fewer "missing pixel" artifacts compared to baseline methods.

Limitations and Future Work

While the paper demonstrates strong results in 2D retinal imaging, the authors do not extensively discuss:

  • Computational Overhead: The use of Monte Carlo Dropout (requiring 10 stochastic forward passes) and prototype calculation adds computational cost during the adaptation phase.
  • 3D Generalization: The current work focuses on 2D fundus images; it remains to be seen how well these specific weighting maps (like the 2D Inverted-Gaussian) translate to complex 3D medical volumes like CT or MRI.
  • Hyperparameter Sensitivity: While the authors provide an ablation study, the framework relies on several thresholds ($\eta, \gamma, \beta$) that might require tuning for significantly different medical tasks.

Summary updated

Abstract

Medical image segmentation models face severe performance drops under domain shifts, especially when data sharing constraints prevent access to source images. We present a novel Uncertainty-aware Progressive Pseudo-label Denoising (UP2D) framework for source-free domain adaptation (SFDA), designed to mitigate noisy pseudo-labels and class imbalance during adaptation. UP2D integrates three key components: (i) a Refined Prototype Filtering module that suppresses uninformative regions and constructs reliable class prototypes to denoise pseudo-labels, (ii) an Uncertainty-Guided EMA (UG-EMA) strategy that selectively updates the teacher model based on spatially weighted boundary uncertainty, and (iii) a quantile-based entropy minimization scheme that focuses learning on ambiguous regions while avoiding overconfidence on easy pixels. This single-stage student-teacher framework progressively improves pseudo-label quality and reduces confirmation bias. Extensive experiments on three challenging retinal fundus benchmarks demonstrate that UP2D achieves state-of-the-art performance across both standard and open-domain settings, outperforming prior UDA and SFDA approaches while maintaining superior boundary precision.
BibTeX
@misc{buitran2025up,
  author = {Quang-Khai Bui-Tran and Thanh-Huy Nguyen and Ho Dung Manh and Thinh B. Lam and Vi Vu and Hoang-Thien Nguyen and Phat K. Huynh and Ulaş Bağcı},
  title = {{UP2D: Uncertainty-aware Progressive Pseudo-label Denoising for Source-Free Domain Adaptive Medical Image Segmentation}},
  year = {2025},
  eprint = {2510.26826},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.26826},
  howpublished = {arXiv},
}
Figure