SABPI-Net: A Structure-Aware Bidirectional Proxy Interaction Network for Infantile Retinal Disease Diagnosis

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Delayed diagnosis of infant retinal diseases leads to irreversible vision loss.
• Automated diagnosis is challenged by subtle lesions, diverse phenotypes, imaging variations, and imbalanced data.
• Existing foundation models struggle with these specific challenges in infant retinal imaging.

Purpose of the Study:

• To develop an advanced AI model for accurate and robust automated diagnosis of infant retinal diseases.
• To enhance the perception of fine retinal structures and ensure consistent performance across diverse imaging conditions.
• To improve diagnostic sensitivity for subtle and less common infant retinal diseases.

Main Methods:

• Proposed the structure-aware bidirectional proxy interaction network (SABPI-Net) within a universal learning framework.
• Incorporated a high-frequency mapping branch with a proxy interaction attention module for detailed structure perception.
• Utilized domain-agnostic embedding space self-matching and a tail-aware feature fusion strategy for fine-tuning.

Main Results:

• SABPI-Net achieved superior performance on three infant retinal disease classification tasks using the largest clinical dataset to date.
• Achieved 95.32% accuracy on mainstream clinical tasks, 73.58% on ROP five-stage classification, and 84.25% on multi-disease classification.
• Demonstrated significant improvements over 13 state-of-the-art methods, with accuracy gains of up to 4.71%.

Conclusions:

• SABPI-Net effectively addresses the challenges in automated infant retinal disease diagnosis.
• The model shows superior diagnostic accuracy and robustness compared to existing methods.
• This advancement holds significant potential for improving early detection and treatment of infant retinal diseases.