Generative Models for Medical Image Creation and Translation: A Scoping Review.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Generative models are crucial for advancing medical imaging.
• Applications include generating new images and translating between modalities.
• Existing reviews lack a comprehensive analysis of diverse generative models and their specific medical imaging tasks.

Purpose of the Study:

• To provide a scholarly review of generative models in medical image creation and translation.
• To analyze the strengths and weaknesses of Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models (DMs).
• To classify and examine tasks within medical image creation and translation, visualizing cross-modality translation trends.

Main Methods:

• Literature review and analysis of generative models (VAEs, GANs, DMs).
• Classification of studies based on downstream tasks (creation) and target modalities (translation).
• Visualization of medical image translation across modalities (MRI, CT, X-ray, PET, ultrasound) and MRI contrast mechanisms using chord diagrams.

Main Results:

• Generative models offer diverse capabilities for medical image synthesis and cross-modality mapping.
• Analysis reveals specific advantages and limitations for VAEs, GANs, and DMs in medical imaging contexts.
• Chord diagrams illustrate the landscape of current research in medical image translation, highlighting key modalities and translation directions.

Conclusions:

• Generative models are powerful tools for medical image creation and translation.
• Understanding model-specific merits and demerits is key for effective application.
• Future research should focus on prospective avenues and implementation guidelines for enhanced medical imaging solutions.