ADAM-Net: Anatomy-Guided Attentive Unsupervised Domain Adaptation for Joint MG Segmentation and MGD Grading.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Meibomian gland dysfunction (MGD) is a primary cause of dry eye disease.
• Deep learning (DL) has improved MGD assessment but struggles with domain shift across different imaging devices.
• Existing DL methods often handle MG segmentation and MGD classification as separate tasks.

Purpose of the Study:

• To develop a robust, multi-task deep learning framework for automated MGD assessment.
• To address the challenge of domain shift in multi-center medical imaging datasets.
• To jointly model meibomian gland segmentation and MGD classification.

Main Methods:

• Proposed ADAM-Net, an attention-guided unsupervised domain adaptation multi-task framework.
• Introduced structure-aware multi-task learning and anatomy-guided attention mechanisms.
• Evaluated cross-domain performance on MGD-1K→{K5M, CR-2, LV II} datasets.

Main Results:

• ADAM-Net achieved classification accuracies of 77.93%, 74.86%, and 81.77% on target domains.
• Significantly outperformed mainstream unsupervised domain adaptation (UDA) methods.
• Demonstrated robust discriminative capability and cross-domain feature alignment, validated by F1-score, MCC-score, and t-SNE visualizations.

Conclusions:

• ADAM-Net provides an effective and robust solution for automated MGD assessment in multi-center scenarios.
• The framework exhibits strong interpretability and addresses domain shift challenges.
• Jointly modeling segmentation and classification enhances performance and robustness.