Fundus Image Enhancement Method Based on CycleGAN.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• High-quality fundus images are crucial for diagnosing eye conditions like diabetic retinopathy.
• Acquiring paired medical images for training deep learning models is challenging and costly.
• Existing image enhancement methods often struggle with preserving texture and detail, especially with unpaired datasets.

Purpose of the Study:

• To develop an unsupervised retinal image enhancement method for migrating low-quality fundus images to high-quality ones.
• To improve the performance of CycleGAN by incorporating the Convolutional Block Attention Module (CBAM) for better detail preservation.
• To validate the effectiveness of the proposed Cycle-CBAM method in enhancing image quality and its impact on diabetic retinopathy classification.

Main Methods:

• Proposed Cycle-CBAM, an enhancement of CycleGAN utilizing the Convolutional Block Attention Module (CBAM).
• Employed an unsupervised learning approach, eliminating the need for paired training datasets.
• Integrated a diabetic retinopathy (DR) classification module to quantitatively assess the impact of image enhancement on diagnostic performance.

Main Results:

• Cycle-CBAM demonstrated superior performance compared to standard CycleGAN in both quantitative and qualitative evaluations.
• The integration of CBAM effectively addressed the degeneration of texture and detail issues.
• Enhanced fundus images showed improved potential for accurate diabetic retinopathy level assessment.

Conclusions:

• The proposed Cycle-CBAM method offers an effective solution for unsupervised retinal image enhancement.
• Incorporating attention mechanisms like CBAM significantly boosts the performance of generative adversarial networks for medical imaging.
• This approach holds promise for improving the accuracy and efficiency of diagnosing eye diseases from fundus images.