WGAN-GP for Synthetic Retinal Image Generation: Enhancing Sensor-Based Medical Imaging for Classification Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Data scarcity is a major challenge in medical image classification.
• Accurate synthetic image generation is vital for improving diagnostic models.
• Diabetic retinopathy classification requires high-quality, diverse datasets.

Purpose of the Study:

• To develop a novel method for generating high-quality synthetic medical images for diabetic retinopathy classification.
• To enhance training datasets using realistic retinal images with preserved pathological features.
• To address data scarcity in sensor-derived medical imaging.

Main Methods:

• Utilized a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP).
• Incorporated nearest-neighbor interpolation for image generation.
• Evaluated performance on multiple retinal image datasets (Retinal-Lesions, FGADR, IDRiD, Kaggle).

Main Results:

• Achieved superior performance compared to traditional generative models (e.g., conditional GANs, PathoGAN).
• Obtained excellent metrics on the Kaggle dataset: FID of 15.21, MSE of 0.002025, SSIM of 0.89.
• Expert evaluation showed only 56.66% of synthetic images were distinguishable from real ones, indicating high fidelity.

Conclusions:

• The proposed WGAN-GP based method effectively generates realistic synthetic retinal images.
• This approach enhances medical image classification by providing high-fidelity, diverse training data.
• The method shows significant potential for improving diabetic retinopathy diagnosis and other medical imaging tasks.