Detection of exudates in fundus photographs with imbalanced learning using conditional generative adversarial network

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision for Disease Detection

Background:

• Diabetic retinopathy (DR) is a primary cause of preventable blindness globally.
• Early diagnosis of DR relies on identifying retinal exudates, which are often missed due to limited labeled data and imbalanced datasets in current deep learning models.
• Existing deep convolutional neural networks (DCNNs) face challenges with small sample sizes and class imbalance, hindering accurate exudate detection.

Purpose of the Study:

• To develop an effective deep learning model for early diabetic retinopathy diagnosis by detecting retinal exudates.
• To address the challenges of limited labeled medical data and severe class imbalance in exudate detection datasets.
• To improve the robustness and generalization capabilities of DCNNs for medical image analysis.

Main Methods:

• An ensemble convolutional neural network (MU-net) based on the U-net architecture was developed to handle small sample sets.
• Conditional generative adversarial networks (cGAN) were employed for data augmentation, specifically generating label-preserving minority class data (exudates).
• The proposed network was trained on one dataset and validated across three independent public datasets (DiaReTDB1, HEI-MED, MESSIDOR).

Main Results:

• The integration of cGAN for data augmentation significantly enhanced model performance across all tested datasets.
• At the lesion level, F1-scores improved with cGAN, reaching up to 94.34%, compared to 90.58% without cGAN.
• At the image level, accuracy increased substantially with cGAN, achieving up to 95.45% compared to 86.42% without cGAN, demonstrating improved generalization.

Conclusions:

• The combination of MU-net and cGAN effectively addresses data scarcity and imbalance issues in diabetic retinopathy exudate detection.
• Data augmentation using cGAN significantly boosts the accuracy and robustness of deep learning models for early DR diagnosis.
• This approach shows strong potential for improving automated screening and early intervention of diabetic retinopathy, thereby reducing blindness.