Abstract
Existing unsupervised domain adaptation (UDA) methods based on adversarial learning have achieved significant performance in medical imaging tasks. However, these methods often focus on global distribution adaptation, ignoring distribution constraints at the category level, which can lead to sub-optimal performance in segmenting fine-grained anatomical structures. This paper presents a UDA framework based on category-level regularization (UDA_CLR) that regularizes category distributions from three perspectives: inter-domain alignment, source intra-domain discriminative learning, and target intra-domain consistency. Extensive experiments on public fundus datasets demonstrate that this approach significantly outperforms state-of-the-art algorithms, closing the gap between unsupervised adaptation and fully supervised performance.
Introduction
In clinical settings, deep learning models for medical image segmentation often encounter the "domain shift" problem. A model trained on a source dataset (e.g., images from one hospital or camera brand) frequently performs poorly on a target dataset from a different environment due to variations in lighting, contrast, and patient demographics.
While Unsupervised Domain Adaptation (UDA) aims to bridge this gap without requiring new labels for the target domain, traditional methods typically rely on global adversarial alignment. This approach attempts to make the overall feature distributions of the source and target domains indistinguishable. However, in tasks like fundus image segmentation—where the goal is to precisely delineate the optic disc and optic cup—global alignment often fails to preserve the distinct boundaries between categories. This paper introduces category-level regularization to ensure that specific anatomical structures remain well-separated and accurately mapped across domains.
Methodology
The proposed UDA_CLR framework addresses domain adaptation through three distinct regularization strategies:
1. Inter-domain Category Regularization
To align features across domains, the framework utilizes "prototypes," which represent the average feature embedding for each category (e.g., optic cup, optic disc, and background). The core challenge is the lack of ground-truth labels in the target domain. To solve this, the authors implement an Uncertainty-Guided Noise-Aware (UGNA) module. By using Monte Carlo dropout, the model estimates the uncertainty of its own predictions on the target data. Only high-confidence "pseudo-labels" are used to calculate target prototypes, preventing noisy or incorrect predictions from degrading the alignment process.
2. Source Intra-domain Regularization
On the labeled source dataset, the model employs a prototype-guided discriminative loss. This loss function acts as a clustering mechanism: it pulls pixel features closer to their corresponding category prototype while pushing them away from prototypes of different categories. This ensures that the model learns highly discriminative and compact feature representations, which are more robust when transferred to a new domain.
3. Target Intra-domain Regularization
To improve stability in the target domain, the framework uses augmented consistency category regularization. The model is required to produce consistent segmentation maps for a target image and its augmented version (e.g., an image modified by noise or color jittering). This forces the network to focus on underlying semantic shapes and structures rather than superficial pixel-level variations.
Experimental Results
The framework was evaluated using the REFUGE dataset as the source domain and two distinct target domains: RIM-ONE-r3 and Drishti-GS. These datasets are benchmarks for glaucoma screening, focusing on the segmentation of the optic disc and cup.
Performance Gains
The UDA_CLR framework established new state-of-the-art (SOTA) results:
- Superiority over SOTA: The method outperformed the previous leading approach (BEAL) by approximately 3.1% and 3.0% in Dice coefficient for optic cup segmentation on the RIM-ONE-r3 and Drishti-GS datasets, respectively.
- Proximity to Oracle: On the Drishti-GS dataset, the model achieved a Dice score of 0.966 for the optic disc, nearing the "Oracle" performance (0.974) achieved by models trained with full supervision on the target data.
Qualitative analysis showed that the proposed method is particularly effective at maintaining smooth and accurate boundaries in regions obscured by blood vessels or characterized by low contrast, where previous global-alignment methods often produced fragmented results.
Discussion and Conclusion
The success of UDA_CLR highlights that global distribution alignment is insufficient for high-precision medical tasks. By focusing on category-level constraints, the model can better preserve the structural integrity of anatomical features across different imaging environments. The use of uncertainty estimation (UGNA) is a critical component, as it allows the model to safely leverage its own predictions on unlabeled data.
While the framework shows great promise, future work may address the computational cost associated with Monte Carlo dropout and explore the generalizability of this category-level approach to 3D medical imaging modalities such as CT or MRI scans. Overall, UDA_CLR represents a significant step toward deploying robust, high-performance AI tools in diverse clinical settings without the need for exhaustive manual labeling.
Summary updated