CycleGAN with Dynamic Criterion for Malaria Blood Cell Image Synthetization.

Area of Science:

• Medical Imaging
• Computational Biology
• Machine Learning

Background:

• Malaria diagnosis relies on accurate blood cell imaging.
• Acquiring diverse, well-annotated malaria-infected blood cell images is challenging.
• Existing generative models may lack sufficient quality and diversity for augmentation.

Purpose of the Study:

• To develop an advanced generative model for synthesizing malaria-infected blood cell images.
• To improve the diversity and quality of synthetic medical images for deep learning applications.
• To enhance the robustness of automated medical image analysis.

Main Methods:

• Implementation of a Cycle-Consistent Adversarial Network (Cycle GAN) with a dynamic criterion.
• Synthesis of blood cells parasitized by malaria plasmodia.
• Quantitative evaluation using a pre-trained classifier and Frechet Inception Distance (FID).

Main Results:

• The enhanced Cycle GAN achieved 100% correct classification of synthetic images.
• Synthetic images showed significantly improved quality and diversity compared to a standard Cycle GAN (76.6% classification) and a Variational Autoencoder (VAE) model (FID=0.0043 vs 0.0085).
• The model generated high-quality malaria-infected blood cell images with good diversity.

Conclusions:

• The proposed Cycle GAN with dynamic criterion is effective for generating high-quality, diverse malaria-infected blood cell images.
• This method offers a valuable image augmentation technique, particularly when annotated data is scarce.
• The approach improves the robustness of deep neural networks in medical image processing.