Unsupervised Domain Adaptation Based Image Synthesis and Feature Alignment for Joint Optic Disc and Cup Segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Ophthalmology

Background:

• Neural networks trained on fundus images often fail on new datasets due to device discrepancies.
• Unsupervised domain adaptation is crucial for generalizing models to unseen data distributions.
• Accurate segmentation of optic disc and cup is vital for diagnosing glaucoma and other eye conditions.

Purpose of the Study:

• To propose an unsupervised domain adaptation method for optic disc and cup segmentation in fundus images.
• To address the domain shift problem caused by variations in fundus image collection devices.
• To improve the generalizability and robustness of segmentation models across different datasets.

Main Methods:

• Developed an Image Synthesis and Feature Alignment (ISFA) method incorporating Generative Adversarial Network (GAN)-based image synthesis.
• Utilized content and style feature alignment (CSFA) to ensure feature consistency between source, synthetic, and target images.
• Employed adversarial learning for output-level feature alignment (OLFA) and an edge attention module (EAM) for enhanced boundary representation.

Main Results:

• Achieved approximately 3% improvement in Dice score for optic cup segmentation on the Drishti-GS dataset compared to the next best method.
• Demonstrated the method's robustness for small dataset domain adaptation.
• Experimental results validated the effectiveness of the proposed ISFA method.

Conclusions:

• The proposed ISFA method effectively alleviates domain shift in fundus image segmentation.
• The combination of image synthesis and feature alignment enhances segmentation performance across diverse datasets.
• This approach offers a promising solution for developing robust and generalizable fundus image analysis tools.