3C-GAN: class-consistent CycleGAN for malaria domain adaptation model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Unpaired domain translation models like CycleGAN are used for medical image synthesis.
• CycleGAN can hallucinate features, potentially altering the original image class and leading to misdiagnosis.
• This is particularly problematic in datasets like malaria images, where features (e.g., parasites) can be unintentionally added or removed.

Purpose of the Study:

• To modify CycleGAN's distribution matching loss to eliminate feature hallucination.
• To ensure domain translation in medical images preserves the original class labels.
• To enhance the safety and reliability of unsupervised generative adversarial networks (GANs) for clinical applications.

Main Methods:

• Introduction of a modified distribution matching loss function for CycleGAN.
• Application of the modified CycleGAN to a malaria image dataset for domain translation.
• Experimental evaluation comparing the modified approach against the classic CycleGAN.

Main Results:

• The modified loss function significantly reduced feature hallucination in synthesized malaria images.
• Original class labels were preserved during domain translation, preventing misclassification.
• Experimental results demonstrated superior performance compared to the baseline CycleGAN.

Conclusions:

• The proposed modified loss effectively eliminates feature hallucination in unpaired domain translation of medical images.
• This approach ensures the integrity of class labels, making GANs safer for clinical use.
• The method holds promise for advancing unsupervised, clinically safe GAN development.