Unsupervised Domain Adaptation Fundus Image Segmentation via Multi-Scale Adaptive Adversarial Learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate segmentation of the optic disc (OD) and optic cup (OC) is vital for early glaucoma detection.
• Deep learning models face challenges in clinical application due to domain shifts in fundus images from different institutions.

Purpose of the Study:

• To introduce an unsupervised domain adaptation technique, Multi-scale Adaptive Adversarial Learning (MAAL), to improve OD and OC segmentation.
• To enhance the generalizability and robustness of segmentation models across diverse fundus image datasets.

Main Methods:

• MAAL integrates three components: Multi-scale Wasserstein Patch Discriminator (MWPD) for domain-specific feature extraction, Adaptive Weighted Domain Constraint (AWDC) for adaptive feature weighting, and Pixel-level Feature Enhancement (PFE) for integrating shallow and deep features.
• The method utilizes unsupervised domain adaptation to bridge the gap between different data domains without requiring labeled data from the target domain.

Main Results:

• The MAAL method demonstrated significant improvement in mitigating model degradation caused by domain shifts.
• Experiments on two public fundus image databases showed that MAAL outperforms current state-of-the-art methods in both OD and OC segmentation.
• The proposed technique effectively preserves domain-invariant information and enhances feature representation.

Conclusions:

• MAAL offers an effective solution for unsupervised domain adaptation in medical image segmentation, specifically for OD and OC.
• The developed technique shows great promise for improving the clinical applicability of deep learning models in ophthalmology.
• This approach enhances segmentation accuracy and robustness, contributing to more reliable glaucoma diagnosis.