Image Translation by Ad CycleGAN for COVID-19 X-Ray Images: A New Approach for Controllable GAN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) are powerful tools for image synthesis.
• Accurate image translation is crucial for medical diagnostics, especially for conditions like COVID-19.
• Class imbalance in medical datasets poses a challenge for AI models.

Purpose of the Study:

• To introduce a novel adaptive Cycle-consistent Generative Adversarial Network (Ad CycleGAN) for image translation.
• To enhance the control and accuracy of generating synthetic chest X-ray images between normal and COVID-19 positive cases.
• To evaluate the performance of Ad CycleGAN against the conventional Cycle GAN.

Main Methods:

• Developed Ad CycleGAN by integrating an independent pre-trained criterion into the Cycle GAN architecture.
• Performed image translation between normal and COVID-19 positive chest X-ray images.
• Quantitatively evaluated image quality using metrics like MSE, RMSE, PSNR, UIQI, VIF, FID, and translation accuracy.

Main Results:

• Both Cycle GAN and Ad CycleGAN produced synthetic images with lower MSE/RMSE and higher PSNR/UIQI/VIF for homogenous translation (Y → Y) compared to heterogeneous translation (X → Y).
• Ad CycleGAN demonstrated a significantly higher Frechet Inception Distance (FID) score than Cycle GAN for heterogeneous translation (p < 0.01).
• Ad CycleGAN achieved higher translation accuracy when converting normal images to COVID-19 positive images (p < 0.01).

Conclusions:

• Ad CycleGAN, with its independent criterion, significantly improves the accuracy of GAN-based image translation.
• The enhanced architecture offers greater control over image synthesis.
• This model can effectively help mitigate the class imbalance problem in medical AI applications.