Introduction
In medical imaging, deep learning models often struggle with Domain Shift. A model trained on fundus images from one hospital (the source) frequently performs poorly when deployed at another site (the target) due to variations in camera hardware, lighting conditions, or patient demographics. While Unsupervised Domain Adaptation (UDA) typically addresses this by retraining on both source and target data, strict privacy regulations often prevent the sharing of sensitive patient data.
This has led to the development of Source-Free Unsupervised Domain Adaptation (SF-UDA), where a model must adapt to a new target site using only the pre-trained source model and unlabeled target images. A common challenge in SF-UDA is the generation of noisy "pseudo-labels"—the model's initial guesses for the target data. These labels often lack context consistency, resulting in segmented anatomical structures with jagged edges or irregular holes. To address this, the authors propose the Context-aware Pseudo-label Refinement (CPR) framework.
The CPR Framework
The CPR framework operates on the insight that while a model's final predictions might be noisy under domain shift, its internal feature representations still tend to cluster similar pixels together. The method proceeds in two primary stages:
1. Context-Similarity Learning
The researchers observed that feature clusters remain relatively robust even when the final segmentation map fails. By freezing the source model's encoder and adding a lightweight "similarity head," the framework learns to calculate the mathematical relationship between pixels. If two pixels are close in the feature space, they are likely part of the same anatomical structure. This stage establishes a foundation for spatial consistency by learning how pixels should relate to their neighbors.
2. Pseudo-Label Refinement and Calibration
Once similarities are learned, the framework cleans the noisy pseudo-labels through three sub-steps:
- Revision: Using the learned similarities, the model performs "probability diffusion." A pixel’s label is updated based on the consensus of its similar neighbors, effectively smoothing out irregular errors and filling in gaps.
- Calibration: Refinement can sometimes inadvertently lower the model's confidence in correct predictions. The calibration step "boosts" these probabilities, ensuring that the most confident pixel in a region remains near a probability of 1.
- Denoising: Finally, a two-level filter (pixel-level and class-level) is applied to discard any remaining low-confidence labels, leaving only high-quality data to guide the final adaptation process.
Experimental Results
The CPR method was evaluated on three major fundus image datasets used for glaucoma screening: Drishti-GS, RIM-ONE-r3, and REFUGE. The experiments focused on cross-domain scenarios, such as adapting a model trained on Drishti-GS to the RIM-ONE-r3 dataset.
Key findings include:
- State-of-the-Art Performance: CPR outperformed existing SF-UDA methods like DPL and U-D4R. On the Drishti-GS to RIM-ONE-r3 task, CPR achieved an average Dice score of 85.03%, a nearly 2% improvement over the previous best method.
- Boundary Precision: The method significantly reduced the Average Surface Distance (ASD), indicating that the boundaries of the segmented optic cup and disc were much closer to the expert "ground truth."
- Ablation Studies: The importance of the calibration step was highlighted when its removal caused the optic cup accuracy to drop from 72% to 58%, proving that smoothing labels alone is insufficient without maintaining model confidence.
Discussion and Limitations
CPR offers a practical path for clinical AI deployment by reducing the memory and time required for training through encoder freezing. However, some challenges remain. The method is sensitive to hyperparameters like the search radius for neighbor similarity and the number of refinement iterations. Additionally, while the local search radius saves computational power, scaling this approach to high-resolution 3D volumes (such as CT or MRI scans) may require further optimization. Finally, the success of CPR depends on the source model's ability to produce meaningful feature clusters; if the domain shift is too extreme, even the underlying features may fail to provide a reliable basis for refinement.
Conclusion
The Context-aware Pseudo-label Refinement (CPR) framework provides a robust solution for source-free domain adaptation in medical imaging. By leveraging the inherent clustering of features and enforcing local context consistency, CPR effectively transforms noisy model guesses into high-quality training signals, paving the way for more reliable and privacy-compliant medical AI.
Summary updated