A Knowledge-Guided Bi-Modal Network for the Classification of Anterior Chamber Angle Images.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a primary cause of irreversible blindness worldwide.
• Anterior Chamber Angle (ACA) evaluation is crucial for glaucoma prognosis and treatment.
• Current ACA evaluation methods are inefficient, relying on labor-intensive expert judgment.

Purpose of the Study:

• To develop an automated ACA classification method using machine learning.
• To address the limitations of image-only deep learning models for ACA evaluation.
• To propose a novel end-to-end knowledge-guided bi-modal network (KGNet) for improved ACA assessment.

Main Methods:

• KGNet integrates two data modalities: textual domain knowledge and medical images.
• Textual knowledge is refined into descriptions to enhance feature diversity.
• A supervised loss incorporates domain knowledge for modality-specific representations, and a fusion module uses knowledge-guided learning for bimodal correlation analysis.

Main Results:

• The proposed KGNet method outperforms state-of-the-art deep learning models on ACA datasets.
• The approach demonstrates significant potential for medical interest in eye image evaluation.
• KGNet enhances interpretability by aligning visual representations with structured clinical knowledge.

Conclusions:

• Automated ACA classification using KGNet offers a more efficient and accurate alternative to traditional methods.
• The integration of textual knowledge with imaging data improves classification performance and clinical relevance.
• KGNet provides clinically grounded explanations, advancing the interpretability of AI in ophthalmology.