Synergic Adversarial Label Learning for Grading Retinal Diseases via Knowledge Distillation and Multi-Task Learning.

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Automated retinal image classification is crucial but limited by expensive expert annotations and scarce data for diseases like diabetic retinopathy (DR) and age-related macular degeneration (AMD).
• Existing algorithms often train models independently, failing to utilize shared pathological features (e.g., hemorrhages, exudation) present in multiple retinal diseases.
• The scarcity of annotated data and the expense of expert labeling present significant challenges in developing robust diagnostic tools for retinal conditions.

Purpose of the Study:

• To introduce a novel method, Synergic Adversarial Label Learning (SALL), for enhanced automated retinal disease classification.
• To leverage multi-task learning principles by integrating relevant disease labels as auxiliary signals in both semantic and feature spaces.
• To improve the accuracy, reliability, and interpretability of AI models for grading retinal diseases like DR and AMD.

Main Methods:

• Developed Synergic Adversarial Label Learning (SALL), a collaborative training approach utilizing knowledge distillation.
• Integrated shared features and relevant disease labels from conditions like DR and AMD as additional signals.
• Employed adversarial learning within a multi-task framework to enhance model robustness and generalization.

Main Results:

• Achieved significant accuracy improvements in DR and AMD fundus image classification: 5.91% and 3.69% respectively.
• Demonstrated the effectiveness of SALL in enhancing model reliability and interpretability for medical imaging applications.
• Validated the collaborative training approach's superiority over independent disease model training.

Conclusions:

• SALL offers a powerful framework for improving automated retinal disease classification by synergistically learning from multiple related conditions.
• The method addresses data scarcity and annotation cost challenges by effectively utilizing shared pathological information.
• SALL enhances the clinical utility of AI in ophthalmology through improved accuracy, reliability, and interpretability.