A deep learning-based framework for retinal fundus image enhancement.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Low-quality fundus images with complex degradation pose challenges for accurate clinical diagnosis and can necessitate costly patient re-examinations.
• Effective enhancement of these images is crucial for improving diagnostic workflows and patient outcomes.

Purpose of the Study:

• To develop an automatic framework for enhancing low-quality fundus macular images and removing complex degradation.
• To improve the quality of retinal fundus images for better clinical interpretation.

Main Methods:

• A deep learning-based model using a customized convolutional neural network (CNN) architecture was proposed.
• A dataset of 1068 high-quality (HQ) and low-quality (LQ) fundus images was collected and used for training.
• Data augmentation techniques were employed to simulate various aspects of retinal image degradation.

Main Results:

• The proposed model significantly increased Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) compared to original LQ images.
• Performance metrics showed significant improvements over existing state-of-the-art methods (P < 0.05).
• The proportion of ungradable fundus photographs decreased from 42.6% to 26.4% (P = 0.012).

Conclusions:

• The developed enhancement process significantly improves the quality of LQ fundus images affected by complex degradation.
• The customized CNN model demonstrated superior performance compared to current state-of-the-art methods.
• This framework has the potential for clinical impact by reducing re-examinations and enhancing diagnostic accuracy.