Unsupervised Domain Adaptive Fundus Image Segmentation with Category-Level Regularization

Wei Feng, Lin Wang, Lie Ju, Xin Zhao, Xin Wang, Xiaoyu Shi, Zongyuan Ge

TL;DR

This paper introduces a new framework for medical image segmentation that works across different datasets without needing new labels. By focusing on "category-level regularization"—ensuring the model distinguishes between specific anatomical structures like the optic disc and cup even when the image quality changes—the authors significantly outperform previous state-of-the-art methods in eye disease screening.

What Problem Does This Paper Address?

In medical AI, a model trained on images from one hospital (the "source domain") often fails when deployed at another (the "target domain"). This is known as domain shift, caused by different camera brands, lighting conditions, or patient demographics.

While researchers use Unsupervised Domain Adaptation (UDA) to fix this, most current methods focus on "global" alignment—trying to make the overall distribution of the two datasets look similar. The authors argue this is insufficient for medical tasks. If the model only looks at the big picture, it might lose the fine-grained boundaries between similar-looking tissues (like the optic cup vs. the optic disc), especially in low-contrast regions. The challenge is: how do you enforce these category-level constraints on a target dataset where you have zero ground-truth labels?

Methods

The authors propose a framework called UDA_CLR (Category-level Regularization). It tackles the problem from three angles:

  1. Inter-domain Alignment (The Bridge): They use "prototypes"—essentially the average feature representation of a category (e.g., "this is what an optic cup looks like"). They align these prototypes between the source and target domains. To handle the lack of labels in the target domain, they use an Uncertainty-Guided Noise-Aware (UGNA) module. This uses Monte Carlo dropout (running the model multiple times with random variations) to filter out "unreliable" guesses before they can mess up the alignment.
  2. Source Intra-domain Regularization (The Foundation): On the labeled source data, they use a prototype-guided discriminative loss. This acts like a "pull-push" mechanism: it pulls pixel features toward their correct category prototype and pushes them away from different ones, creating very distinct clusters in the feature space.
  3. Target Intra-domain Regularization (The Consistency): On the unlabeled target data, they use augmented consistency. They take a target image, perturb it (blur it or change the colors), and force the model to produce the same segmentation map for both the original and the "messy" version. This encourages the model to focus on semantic shapes rather than superficial pixel values.

Datasets

The researchers evaluated their model on three well-known fundus (retinal) image datasets:

  • REFUGE: Used as the source domain (400 annotated images).
  • RIM-ONE-r3: Used as a target domain (99 training / 60 testing images).
  • Drishti-GS: Used as a second target domain (50 training / 51 testing images).

These datasets are standard benchmarks for segmenting the optic disc and cup, which are critical for diagnosing glaucoma.

Key Insights and Ideas

  • Global isn't enough: The "aha moment" here is that even if two datasets look similar globally, the "boundaries" between classes can still be mushy. Category-level focus is what provides clinical-grade precision.
  • Trust, but verify: Using pseudo-labels (the model's own guesses) to train a model is risky because errors can compound. The use of uncertainty estimation (UGNA) to "vet" these labels before using them for prototype alignment is a clever way to prevent the "blind leading the blind."
  • Discriminative Source Training: By making the source features more "extreme" (highly separated) through the discriminative loss, the model learns features that are inherently more robust and easier to map to a new domain.

Key Results

The proposed method set a new bar for performance on the RIM-ONE-r3 and Drishti-GS datasets:

  • Superiority over SOTA: It outperformed the previous best method (BEAL) by 3.1% and 3% in Dice coefficient (a measure of overlap) for optic cup segmentation.
  • Closing the Gap: The model's performance is now remarkably close to the "Oracle"—a model trained with full access to the target domain's labels. For example, on Drishti-GS, the model achieved a 0.966 Dice score for the optic disc, compared to the Oracle's 0.974.
  • Visual Accuracy: Qualitative results show the model is much better at handling regions obscured by blood vessels or low contrast, where previous models often produced "jagged" or incorrect boundaries.

Limitations and Future Work

While the results are impressive, the paper leaves a few stones unturned:

  • Computational Overhead: The uncertainty estimation requires multiple forward passes (Monte Carlo dropout), which increases the training time and resource requirements.
  • Generalization: The study focuses specifically on fundus images. While the authors suggest it can be applied to other tasks, its performance on more complex 3D data (like CT or MRI scans) remains to be seen.
  • Hyperparameter Sensitivity: The framework relies on several "balance coefficients" ($\lambda_1$ through $\lambda_4$) and distance margins ($\delta$). Finding the "sweet spot" for these values in a completely new domain without any labels could be challenging in a real-world clinical setting.

Summary updated

Abstract

Existing unsupervised domain adaptation methods based on adversarial learning have achieved good performance in several medical imaging tasks. However, these methods focus only on global distribution adaptation and ignore distribution constraints at the category level, which would lead to sub-optimal adaptation performance. This paper presents an unsupervised domain adaptation framework based on category-level regularization that regularizes the category distribution from three perspectives. Specifically, for inter-domain category regularization, an adaptive prototype alignment module is proposed to align feature prototypes of the same category in the source and target domains. In addition, for intra-domain category regularization, we tailored a regularization technique for the source and target domains, respectively. In the source domain, a prototype-guided discriminative loss is proposed to learn more discriminative feature representations by enforcing intra-class compactness and inter-class separability, and as a complement to traditional supervised loss. In the target domain, an augmented consistency category regularization loss is proposed to force the model to produce consistent predictions for augmented/unaugmented target images, which encourages semantically similar regions to be given the same label. Extensive experiments on two publicly fundus datasets show that the proposed approach significantly outperforms other state-of-the-art comparison algorithms $$^1$$ (Our code is available at https://github.com/fengweie/UDA_CLR .).
BibTeX
@inproceedings{feng2022unsupervised,
  author = {Wei Feng and Lin Wang and Lie Ju and Xin Zhao and Xin Wang and Xiaoyu Shi and Zongyuan Ge},
  title = {{Unsupervised Domain Adaptive Fundus Image Segmentation with Category-Level Regularization}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  url = {https://openreview.net/forum?id=Pek0gA8pfN9},
}

Loading related content...

Figure