Exploiting the Generative Adversarial Network Approach to Create a Synthetic Topography Corneal Image

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Corneal diseases are prevalent eye disorders often diagnosed using automated methods.
• Deep learning (DL) models for medical image analysis require extensive annotated datasets, a significant limitation.
• Existing datasets for corneal disease diagnosis are often imbalanced, affecting classifier performance.

Purpose of the Study:

• To present a method for synthesizing medical images of corneal diseases using conditional generative adversarial networks (CGANs).
• To demonstrate the utility of synthesized images in augmenting medical data for improved clinical decisions.
• To evaluate the impact of data augmentation and balancing on the performance of conventional neural networks (CNNs) for corneal image diagnosis.

Main Methods:

• Corneal topography images from 3448 patients with corneal diseases were collected using a Pentacam device.
• Conditional generative adversarial networks (CGANs) were employed to synthesize new corneal images.
• A resampling approach was used to balance the dataset, and CNNs were trained on both balanced and imbalanced datasets.
• Performance was evaluated using accuracy, precision, and F1-score metrics, with expert evaluation of synthesized images.

Main Results:

• Synthesized medical images were found to be useful for medical diagnosis and severity classification.
• CNNs trained on the balanced dataset, augmented with synthesized images, showed improved performance compared to those trained on the imbalanced dataset.
• Expert evaluation confirmed the clinical utility of generated images for identifying disease stages.

Conclusions:

• Conditional generative adversarial networks (CGANs) offer a viable solution for generating synthetic medical images to address data scarcity in corneal disease diagnosis.
• Data augmentation and balancing techniques significantly improve the performance of deep learning models in diagnosing corneal conditions.
• The synthesized images hold potential for enhancing medical education, clinical decision support, and the development of more robust diagnostic tools.