Evaluating Generative Models in Medical Imaging.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Medical Imaging

Background:

• Data availability is a critical challenge in biomedical informatics.
• Generative models offer potential solutions for data synthesis.
• Evaluating these models is crucial for understanding their utility.

Purpose of the Study:

• To examine state-of-the-art generative models for medical imaging data synthesis.
• To quantitatively evaluate the performance of models like StyleGAN and DDPM.
• To assess their ability to learn data manifolds and generate realistic visible features.

Main Methods:

• Review and analysis of leading generative models (StyleGAN, DDPM) in medical imaging.
• Performance evaluation based on data manifold learning capabilities.
• Assessment of the quality of visible features in synthesized samples.

Main Results:

• Current generative models demonstrate limitations in learning complex data manifolds.
• The visible features of generated medical images require substantial enhancement.
• Existing models fall short of optimal performance based on key evaluation metrics.

Conclusions:

• Generative models have potential but need significant advancement for reliable biomedical data synthesis.
• Further research is necessary to improve manifold learning and feature generation in medical imaging.
• Current models are not yet sufficient for widespread application in addressing data availability challenges.