Convolutional autoencoder joint boundary and mask adversarial learning for fundus image segmentation.

Area of Science:

• Ophthalmology
• Medical Image Analysis
• Computer Vision

Background:

• Accurate segmentation of the optic cup (OC) and optic disc (OD) is crucial for effective glaucoma screening.
• Convolutional neural networks (CNNs) have shown promise in medical image segmentation but often struggle with generalization across different datasets (cross-domain problem).

Purpose of the Study:

• To develop a novel unsupervised domain-adaptive segmentation architecture (CAE-BMAL) for improved optic cup and optic disc segmentation.
• To address the challenge of domain generalization in medical image segmentation for glaucoma screening.

Main Methods:

• Employed a convolutional autoencoder (CAE) to enhance the source domain data, boosting model generalization.
• Introduced an adversarial learning-based boundary discrimination branch to mitigate segmentation errors in complex environments.
• Evaluated the proposed CAE-BMAL architecture on three diverse datasets: Drishti-GS, RIM-ONE-r3, and REFUGE.

Main Results:

• The CAE-BMAL method demonstrated superior performance compared to most state-of-the-art methods in terms of accuracy and generalization.
• Experimental evaluations confirmed the effectiveness of the proposed approach on multiple datasets.
• The cup-to-disk ratio analysis further validated the method's utility in glaucoma discrimination.

Conclusions:

• The proposed CAE-BMAL architecture offers a robust solution for unsupervised domain-adaptive segmentation of optic disc and optic cup.
• This method shows significant potential for enhancing the accuracy and reliability of automated glaucoma screening systems.
• The approach effectively tackles the cross-domain generalization challenge in medical image segmentation.