An adaptive weighted ensemble learning network for diabetic retinopathy classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness globally.
• Automated early DR detection faces challenges due to imbalanced data distributions.
• Optical coherence tomography (OCT) is crucial for visualizing retinal structures in DR diagnosis.

Purpose of the Study:

• To propose an adaptive weighted ensemble learning method for enhanced DR detection using OCT images.
• To address the challenge of imbalanced datasets in automated DR detection.
• To improve the overall performance and reliability of deep learning models for DR screening.

Main Methods:

• Developed an ensemble learning model integrating three advanced deep learning models.
• Implemented a novel decision fusion scheme based on Bayesian theory to dynamically adjust model weights.
• Utilized key evaluation indicators to mitigate the impact of unbalanced data size.

Main Results:

• Achieved a quadratic weighted kappa of 0.8487 and accuracy of 0.9343 on the DRAC2022 dataset.
• Obtained a quadratic weighted kappa of 0.9007 and accuracy of 0.8956 on the APTOS2019 dataset.
• Demonstrated significant enhancement in DR detection performance on OCT images.

Conclusions:

• The proposed adaptive weighted ensemble method effectively improves DR detection accuracy.
• The Bayesian-based fusion scheme successfully alleviates issues caused by imbalanced data.
• This approach shows promise for reliable automated early detection of diabetic retinopathy.