Pix2pix Conditional Generative Adversarial Networks for Scheimpflug Camera Color-Coded Corneal Tomography Image...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Keratoconus diagnosis relies on corneal tomography, but training diagnostic models is hindered by limited datasets.
• Data augmentation is crucial for improving the performance of deep learning models in medical image analysis.

Purpose of the Study:

• To evaluate the pix2pix conditional generative adversarial network (cGAN) for synthesizing realistic Scheimpflug camera corneal tomography images.
• To assess the utility of these synthesized images for augmenting training data for a deep convolutional neural network (DCNN) used in keratoconus classification.

Main Methods:

• Retrospective analysis of 1778 corneal tomography images from 923 patients.
• Training a pix2pix cGAN with original images and evaluating synthesized image quality using metrics like Fréchet inception distance, mean square error, and structural similarity index.
• Comparing DCNN classification performance using original images, traditionally augmented images, and a combination of real and synthesized images.

Main Results:

• The pix2pix cGAN generated plausible synthesized corneal tomography images, validated both subjectively and objectively.
• Training a DCNN with a mix of real and synthesized images resulted in superior classification performance compared to using only original or traditionally augmented images.

Conclusions:

• The pix2pix cGAN effectively synthesizes high-quality corneal tomography images, addressing dataset limitations and class imbalance in training diagnostic models.
• This approach offers a scalable solution for generating synthetic data for computer-aided diagnosis in ophthalmology.