Source-linked AI summary
Boundary and Entropy-driven Adversarial Learning for Fundus Image Segmentation
Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, Pheng-Ann Heng
TL;DR
Domain shift limits the generalization of OD and OC segmentation models across fundus datasets, despite the importance of accurate segmentation for glaucoma screening. BEAL uses adversarial learning to align target-domain boundary predictions and entropy maps with source-domain outputs. On Drishti-GS and RIM-ONE-r3, it outperforms state-of-the-art unsupervised domain adaptation methods, including a 2.3% OC and 3.3% OD DI improvement over pOSAL on RIM-ONE-r3.
Problem
Domain shift hinders deep neural networks’ generalization across fundus datasets, limiting accurate OD and OC segmentation needed for glaucoma screening.
Method
BEAL uses boundary and mask branches plus adversarial learning to align target-domain boundary structure and prediction entropy maps with source-domain outputs.
Results
2.3% OC and 3.3% OD DI improvement over pOSAL on RIM-ONE-r3, with BEAL outperforming state-of-the-art unsupervised domain adaptation methods across Drishti-GS and RIM-ONE-r3.
Takeaways & Limitations
Simultaneously enforcing domain-invariant boundaries and entropy maps produces more accurate boundaries and suppresses uncertain OD and OC predictions across the evaluated target domains.
Takeaways & Limitations
The evaluation uses REFUGE training data as source and Drishti-GS and RIM-ONE-r3 as target domains without target-domain ground-truth supervision.
Abstract
from arXiv · showhide
Accurate segmentation of the optic disc (OD) and cup (OC)in fundus images from different datasets is critical for glaucoma disease screening. The cross-domain discrepancy (domain shift) hinders the generalization of deep neural networks to work on different domain datasets.In this work, we present an unsupervised domain adaptation framework,called Boundary and Entropy-driven Adversarial Learning (BEAL), to improve the OD and OC segmentation performance, especially on the ambiguous boundary regions. In particular, our proposed BEAL frame-work utilizes the adversarial learning to encourage the boundary prediction and mask probability entropy map (uncertainty map) of the target domain to be similar to the source ones, generating more accurate boundaries and suppressing the high uncertainty predictions of OD and OC segmentation. We evaluate the proposed BEAL framework on two public retinal fundus image datasets (Drishti-GS and RIM-ONE-r3), and the experiment results demonstrate that our method outperforms the state-of-the-art unsupervised domain adaptation methods. Codes will be available at https://github.com/EmmaW8/BEAL.
1 Introduction
OD and OC segmentation supports glaucoma screening, but domain shift limits supervised models’ generalization across fundus datasets. BEAL addresses this shift by aligning target-domain boundary structure and prediction certainty with source-domain behavior through adversarial learning.
- Motivation: Domain shift causes supervised deep networks to produce unsatisfactory OD and OC predictions on new fundus datasets.M-Net performs strongly on ORIGA but generalizes poorly to other testing datasets.
- Motivation: Unsupervised domain adaptation reduces reliance on costly and time-consuming target-domain annotations.Prior approaches include image appearance transfer and high-level feature alignment across domains.
- Proposed approach: BEAL aligns boundary predictions and entropy maps between source and target domains using two adversarial learning objectives.Its framework uses a DeepLabv3+ backbone, boundary and mask branches, Shannon entropy, and separate discriminators.
- Proposed approach: Source-domain predictions tend to have structured boundaries and low entropy, whereas target-domain predictions show ambiguous boundaries and noisy high-entropy regions.The method therefore seeks domain-invariant boundary structure and more certain target-domain predictions.
2 Methodology
BEAL combines boundary- and entropy-driven adversarial learning with supervised source-domain segmentation losses to adapt OD and OC predictions across fundus-image domains. Its discriminators align target boundary structures and entropy maps with source-domain patterns, while a lightweight DeepLabv3+-based architecture produces boundary and mask predictions.
- Boundary-driven Adversarial Learning: BEAL uses boundary-driven adversarial learning to make target-domain boundary predictions resemble the more structured source-domain boundaries.A boundary prediction branch and mask prediction branch are added to the segmentation network, while a boundary discriminator distinguishes source from target boundary predictions.
- Entropy-driven Adversarial Learning: Entropy-driven adversarial learning aligns target and source entropy-map distributions to suppress uncertain target-domain segmentation predictions.The entropy map is computed from pixel-wise mask probabilities, and an entropy discriminator is trained alongside the segmentation network.
- Network Architecture and Training Procedure: BEAL adapts DeepLabv3+ by replacing Xception with MobileNetV2 and adding boundary and mask prediction branches after high- and low-level feature concatenation.The backbone modification reduces parameters and accelerates computation; OD and OC probability maps are generated simultaneously, with entropy maps jointly supplied to the discriminator.
- Network Architecture and Training Procedure: The segmentation network combines source-domain mask and boundary losses with target-domain adversarial loss during alternating optimization.The overall objective includes mask prediction loss, boundary regression loss, and a balance-weighted adversarial term; source and target images are used for their respective objectives.
- Evaluation: Evaluation uses REFUGE training data as the source domain and Drishti-GS and RIM-ONE-r3 data as target domains, with OD and OC Dice coefficients as metrics.The comparison includes a no-adaptation baseline, a supervised upper bound, and several unsupervised domain adaptation methods.
3 Experiments and Results
BEAL is evaluated across two target fundus datasets using quantitative, qualitative, and component-wise experiments. The results compare domain adaptation methods and show benefits from boundary and entropy-driven adversarial learning.
- Experimental setup: The REFUGE training dataset serves as source domain, while Drishti-GS and RIM-ONE-r3 provide target-domain training and testing data.Dataset statistics are reported in Table 1.
- Implementation details: The framework is trained directly in PyTorch without a supervised warm-up phase using minibatches of size 8 and separate SGD and Adam optimization.The discriminators use SGD, while the segmentation network uses Adam, over 200 epochs.
- Quantitative evaluation: Dice coefficients (DI) for optic disc and cup segmentation are compared against a baseline, supervised upper bound, and multiple unsupervised adaptation methods.Compared methods include TD-GAN, high-level feature alignment, and output space-based adaptation.
- Quantitative evaluation: 2.3% and 3.3% DI improvements over pOSAL are reported for optic cup and optic disc segmentation, respectively, on RIM-ONE-r3.The comparison demonstrates the reported effectiveness of BEAL on this target dataset.
- Qualitative evaluation: On RIM-ONE-r3, BEAL produces more accurate boundaries and cleaner entropy maps than pOSAL, which struggles in ambiguous regions and generates high entropy values.The figure visualizes disc and cup contours with green and blue lines; entropy values are rescaled to [0,1].
- Ablation study: Ablation results show that boundary constraints improve both datasets, each adversarial component improves OD and OC segmentation, and their combination yields further improvement.The evaluated variants include DeepLabv3+ without a boundary branch, a boundary-equipped baseline, BAL, EAL, and BEAL.
4 Conclusion
The paper concludes that BEAL addresses domain shift by making boundary predictions and prediction entropy maps domain-invariant. It reports improved OD and OC segmentation across two public fundus datasets and suggests broader applicability to unsupervised domain adaptation.
- 4 Conclusion: BEAL simultaneously encourages domain-invariant boundary predictions and entropy maps to generate more accurate and less uncertain OD and OC predictions.The method is proposed for fundus images from different domains.
- 4 Conclusion: BEAL outperforms state-of-the-art methods on the Drishti-GS and RIM-ONE-r3 public fundus segmentation datasets.The conclusion presents this as evidence of the method’s effectiveness across the evaluated datasets.
- 4 Conclusion: The authors state that BEAL could be generalized to other unsupervised domain adaptation problems.This conclusion extends the stated scope beyond optic disc and cup segmentation.