Source-linked AI summary
Source-Free Domain Adaptive Fundus Image Segmentation with Denoised Pseudo-Labeling
Cheng Chen, Quande Liu, Yueming Jin, Qi Dou, Pheng-Ann Heng
TL;DR
The paper studies source-free domain adaptation when only a trained source model and unlabeled target data are available. It adapts the model through denoised pseudo-labeling with uncertainty- and prototype-based filtering, achieving comparable or higher performance than source-dependent UDA methods on cross-domain fundus segmentation.
Problem
Source-free UDA asks how a model can adapt to an unlabeled target domain when source data are inaccessible because of privacy and data-sharing constraints.
Method
The method performs target-domain self-training with pseudo labels, filtering unreliable labels through pixel-level uncertainty estimation and class-level prototype estimation.
Results
90.13% and 96.39% Dice scores for optic disc segmentation were achieved on the two target datasets without source data, exceeding the BEAL method.
Takeaways & Limitations
Denoised pseudo-labeling provides a practical source-free adaptation solution for cross-domain fundus image segmentation without altering source training.
Takeaways & Limitations
The method assumes generally minor domain shift; under more severe cross-modality shift, the source network may largely under-segment target objects.
Abstract
from arXiv · showhide
Domain adaptation typically requires to access source domain data to utilize their distribution information for domain alignment with the target data. However, in many real-world scenarios, the source data may not be accessible during the model adaptation in the target domain due to privacy issue. This paper studies the practical yet challenging source-free unsupervised domain adaptation problem, in which only an existing source model and the unlabeled target data are available for model adaptation. We present a novel denoised pseudo-labeling method for this problem, which effectively makes use of the source model and unlabeled target data to promote model self-adaptation from pseudo labels. Importantly, considering that the pseudo labels generated from source model are inevitably noisy due to domain shift, we further introduce two complementary pixel-level and class-level denoising schemes with uncertainty estimation and prototype estimation to reduce noisy pseudo labels and select reliable ones to enhance the pseudo-labeling efficacy. Experimental results on cross-domain fundus image segmentation show that without using any source images or altering source training, our approach achieves comparable or even higher performance than state-of-the-art source-dependent unsupervised domain adaptation methods.
1 Introduction
The paper addresses source-free unsupervised domain adaptation when source data cannot be accessed, using only a trained source model and unlabeled target data. It proposes denoised pseudo-labeling to support target-domain self-adaptation despite noisy predictions caused by domain shift.
- Problem: Source-free adaptation uses only a well-trained source model and unlabeled target data, avoiding source-data transmission during target-domain adaptation.This setting is motivated by privacy and ethical constraints surrounding medical data sharing.
- Problem: The central challenge is enabling model self-adaptation from target-domain knowledge when source data and source distribution information are unavailable.
- Approach: Denoised pseudo-labeling generates target supervision from the source model without using source-domain data or altering source-domain training.The method builds on self-training with model predictions as pseudo labels.
- Approach: Pixel-level uncertainty estimation identifies unreliable pseudo labels, while class-level prototype estimation flags labels far from their corresponding class prototypes.These complementary schemes are designed to reduce noise and provide more discriminative supervision for adaptation.
- Results: On cross-domain fundus image segmentation, the method achieves comparable or higher performance than state-of-the-art source-dependent UDA methods without source data.
2 Method
The method adapts a source model using unlabeled target images and denoised pseudo labels, without source images. It combines pixel-level uncertainty estimation with class-level prototype estimation to filter unreliable predictions before target-domain self-training.
- 2.1 Pseudo-Labeling for Source-free UDA: Source-free UDA adapts a source model using only unlabeled target data, with pseudo labels generated from source-model predictions for supervised target-domain training.The target model is initialized from the source model and updated over target data.
- 2.1 Pseudo-Labeling for Source-free UDA: Pseudo labels can accumulate segmentation errors under domain shift, so the method filters unreliable labels rather than relying solely on prediction confidence.The paper notes that domain-shifted models can be over-confident on incorrect predictions.
- 2.2 Pixel-level Denoising via Uncertainty Estimation: Pixel-level denoising estimates uncertainty from K stochastic Monte Carlo Dropout predictions and selects a pixel when its uncertainty is below threshold η.The uncertainty map is computed as the standard deviation of the stochastic outputs; selected labels contribute to the loss.
- 2.3 Class-level Denoising via Prototype Estimation: Class-level denoising estimates object and background prototypes from uncertainty-guided features, then rejects labels whose features are closer to the opposite prototype.Prototype distances are computed from interpolated feature maps, with prediction probabilities weighting prototype construction.
- Training Procedure: The final selection mask combines low prediction uncertainty with relative feature proximity to the predicted class prototype before updating the segmentation loss.The denoised pseudo labels are generated from source-model predictions, filtered, and then used to optimize the target model.
3 Experiments
The experiments evaluate source-free adaptation for cross-domain fundus optic disc and cup segmentation against strong domain-adaptation baselines. Results show that denoising improves pseudo-label reliability and adaptation performance, while applying denoising as post-processing can be harmful.
- Experimental setup: The evaluation uses REFUGE as source and RIM-ONE-r3 and Drishti-GS as target datasets for optic disc and cup segmentation.The source domain contains 400 annotated images, while target training/testing splits are 99/60 and 50/51 images.
- Baselines: The comparison includes BEAL, AdvEnt, SRDA, and DAE, which use more source-domain information than the proposed source-free method.BEAL, AdvEnt, SRDA, and DAE either access source data or alter source training, whereas the proposed method does neither.
- Comparison with state-of-the-art: 90.13% and 96.39% Dice scores for optic disc segmentation on the two target datasets exceed the BEAL baseline.The comparison also includes W/o adaptation and Oracle reference settings.
- Ablation study: Pixel-level and class-level denoising each improve Dice and ASD over plain pseudo-labeling, while their combination yields further gains on both target domains.The two schemes identify different subsets of noisy pseudo labels, making their combination complementary.
- Ablation study: Both denoising schemes improve selected pseudo-label accuracy, and their combination increases accuracy for optic disc and cup across both target domains.The schemes are more effective when denoising pseudo labels than when post-processing segmentation results.
- Ablation study: Post-processing with the denoising schemes performs even lower than the W/o adaptation model.Pseudo-label denoising tolerates some noisy-label misidentification, whereas post-processing errors directly reduce segmentation performance.
4 Conclusion
The paper presents a source-free UDA method that performs competitively on cross-domain fundus segmentation without source data or altered source training. Its applicability is strongest under minor domain shifts, while severe cross-modality shifts remain challenging.
- The method achieves comparable or better performance than source-data-dependent approaches on cross-domain fundus image segmentation without accessing source data or altering source training.
- The proposed pseudo-label denoising schemes may also apply to semi-supervised learning.
- The approach assumes a generally minor domain shift under which pseudo labels provide a meaningful supervision base.For severe cross-modality shifts, under-segmentation may leave too few positive-class pseudo labels for effective self-adaptation.