Scholarlink

Purpose <p>Magnetic Resonance Imaging (MRI) is an essential tool in medical diagnostics, providing high-resolution images across various contrasts that highlight different tissue properties. However, acquiring multiple contrast images for a single patient can be time-consuming and costly. Recent advances in deep learning have shown promise in synthesizing cross-contrast MRI images, enabling the generation of missing or unobtainable contrasts from existing data.</p> Methods <p>This review provides a comprehensive overview of 30 carefully selected studies from recent literature, highlighting the latest deep learning approaches applied to cross-contrast MRI image synthesis. Our focus includes Generative Adversarial Networks (GANs), Convolutional Neural Networks (CNNs), Denoising Diffusion Probabilistic Models (DDPMs), and their variants. We summarize key methodologies, evaluate performance based on image quality and clinical utility, and examine the characteristics of the primary datasets used.</p> Results <p>Challenges such as anatomical structure retention and the overall quality of synthesized images are discussed, along with the practical applications of synthetic MRI images in clinical settings. Our findings indicate that GAN- and DDPM-based models currently lead in performance, achieving state-of-the-art results across various metrics.</p> Conclusion <p>Finally, we highlight open areas for future research, particularly the need for models that can better adapt to varying clinical settings and imaging protocols, and for studies with a greater focus on clinical utility of synthesized images. This review aims to bridge the gap between the technical advancements in deep learning and their potential to enhance MRI-based clinical workflows.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A review on cross-contrast MRI image synthesis through deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月24日

查看全文
  • Richard Acs,
  • Hanqi Zhuang

摘要

Purpose

Magnetic Resonance Imaging (MRI) is an essential tool in medical diagnostics, providing high-resolution images across various contrasts that highlight different tissue properties. However, acquiring multiple contrast images for a single patient can be time-consuming and costly. Recent advances in deep learning have shown promise in synthesizing cross-contrast MRI images, enabling the generation of missing or unobtainable contrasts from existing data.

Methods

This review provides a comprehensive overview of 30 carefully selected studies from recent literature, highlighting the latest deep learning approaches applied to cross-contrast MRI image synthesis. Our focus includes Generative Adversarial Networks (GANs), Convolutional Neural Networks (CNNs), Denoising Diffusion Probabilistic Models (DDPMs), and their variants. We summarize key methodologies, evaluate performance based on image quality and clinical utility, and examine the characteristics of the primary datasets used.

Results

Challenges such as anatomical structure retention and the overall quality of synthesized images are discussed, along with the practical applications of synthetic MRI images in clinical settings. Our findings indicate that GAN- and DDPM-based models currently lead in performance, achieving state-of-the-art results across various metrics.

Conclusion

Finally, we highlight open areas for future research, particularly the need for models that can better adapt to varying clinical settings and imaging protocols, and for studies with a greater focus on clinical utility of synthesized images. This review aims to bridge the gap between the technical advancements in deep learning and their potential to enhance MRI-based clinical workflows.