Medical image translation with deep learning: Advances, datasets and perspectives

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Traditional medical image generation often omits patient-specific data, limiting clinical use.
• Medical image translation accurately converts images between modalities, preserving anatomical and cross-modal features.
• This technique offers advantages for model development and clinical practice.

Purpose of the Study:

• To review recent advancements in deep learning-based medical image translation.
• To elaborate on diverse tasks, applications, and fundamental models used in the field.
• To identify future trends, challenges, and research directions.

Main Methods:

• Overview of fundamental models: Convolutional Neural Networks (CNNs), Transformers, and State Space Models (SSMs).
• Exploration of generative models: Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Autoregressive Models (ARs), Diffusion Models, and Flow Models.
• Discussion of evaluation metrics and commonly used datasets for assessing translation quality.

Main Results:

• Detailed review of deep learning techniques applied to medical image translation.
• Analysis of various generative models and their suitability for different translation tasks.
• Identification of key evaluation metrics and datasets crucial for progress in the field.

Conclusions:

• Deep learning-based medical image translation shows significant promise for clinical applications.
• Continued research in advanced models, evaluation metrics, and datasets is essential.
• This review serves as a reference for researchers to drive innovation in medical image translation.