A Systematic Review on Synthetic Medical Images Generation-Recent Trends and Future Opportunities

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• High-quality annotated medical datasets are scarce and expensive due to privacy, cost, and availability.
• Generative AI models offer a solution for synthesizing complex, realistic medical images.

Purpose of the Study:

• To provide a comparative overview of Generative AI models for medical image generation.
• To categorize research into Generative Adversarial Networks (GANs), Diffusion Models, and Autoencoders.
• To analyze Generative Adversarial Network (GAN) loss functions.

Main Methods:

• Systematic literature review following PRISMA 2020 guidelines.
• Screening of studies published between 2021 and 2026 from scientific databases.
• Selection of 120 papers focusing on generative modeling for medical imaging.

Main Results:

• Generative Adversarial Networks (GANs) were most common (36%), achieving low FID scores (15.552-31.349).
• Diffusion Models (26%) demonstrated superior structural fidelity (SSIM up to 0.915).
• Hybrid loss functions improved image quality (PSNR=35.6, SSIM=0.95).

Conclusions:

• Generative Adversarial Networks (GANs) produce realistic textures, while Diffusion Models yield high-fidelity images but demand significant computational resources.
• Challenges include the need for annotated datasets, high-quality training images, and data availability.
• Clinical implications and limitations of Generative AI in medical imaging were discussed.