Ensemble learning for retinal disease recognition under limited resources

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal optical coherence tomography (OCT) imaging is vital for diagnosing posterior ocular segment diseases.
• Automated analysis of OCT images is crucial for clinical decision-making, but deep learning (DL) models require extensive data and computational power.
• Data acquisition challenges (privacy, labeling) and resource limitations hinder DL model development in medical AI.

Purpose of the Study:

• To develop a novel ensemble learning mechanism for retinal disease recognition.
• To address challenges of limited data and computational resources in medical AI.
• To improve the performance of DL models for retinal OCT image analysis.

Main Methods:

• Proposed a novel ensemble learning mechanism leveraging pre-trained models for knowledge transfer to retinal OCT images.
• Developed an approach that requires fewer parameters compared to training DL models from scratch.
• Utilized multiple pre-trained models to create ensemble models for disease recognition.

Main Results:

• The proposed ensemble method demonstrated superior performance over baseline models on sparse labeled data.
• The triple ensemble model achieved an accuracy of 92.06%, outperforming baseline models by 8.27% to 11.14%.
• The triple ensemble model required fewer trainable parameters (3.677M) compared to baseline models trained from scratch.

Conclusions:

• The novel ensemble learning mechanism effectively recognizes retinal diseases with limited data and computational resources.
• This approach offers a robust solution for developing high-performance medical AI, especially in resource-constrained environments.
• The findings highlight the potential of ensemble learning to overcome data and computational limitations in medical image analysis.