Scholarlink

Purpose <p>In medical imaging, MRI and CT are two prominent technologies. MRI provides clear images of soft tissues, while CT is useful for diagnosing bone and joint conditions, as well as detecting cancer, heart disease, and internal injuries. However, CT involves exposure to ionizing radiation, which, although presenting a small risk, can potentially cause cancer or genetic mutations in future generations. In addition, some patients may experience side effects from contrast dyes used in CT imaging. The cross-modality image synthesis from MRI to CT offers a promising solution to the limitations inherent to CT. Cycle consistent adversarial networks(CycleGAN), a high-performing architecture in the unpaired image-to-image synthesis, designed to produce convincing synthetic CT images from MRI images of the same subject without the need for an actual CT scan. However, CycleGAN often struggles to learn richer or more abundant image features and is difficult to capture structural information.</p> Methods <p>To address these issues, we developed a novel capsule cycle generative adversarial network (Capsule-CycleGAN). Capsule-CycleGAN is specifically designed to handle scenarios where paired data is not available. We integrate the capsule architecture into the CycleGAN discriminator, departing from the conventional convolutional neural network approach.</p> Result <p>To boost the capability of capturing relevant features, we refined the entire objective function by integrating margin loss, identity loss, cycle consistency, and adversarial loss, and we have achieved significant improvements in quantitative evaluation measures.</p> Conclusion <p>The effectiveness of our method was validated through comprehensive evaluations, demonstrating superior performance compared to other leading approaches in the synthesizing of high-quality CT images from MRI input.</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

Unsupervised MRI to CT image synthesis using Capsule-CycleGAN

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • M. S. Meharban,
  • M. K. Sabu,
  • T. Santhanakrishnan

摘要

Purpose

In medical imaging, MRI and CT are two prominent technologies. MRI provides clear images of soft tissues, while CT is useful for diagnosing bone and joint conditions, as well as detecting cancer, heart disease, and internal injuries. However, CT involves exposure to ionizing radiation, which, although presenting a small risk, can potentially cause cancer or genetic mutations in future generations. In addition, some patients may experience side effects from contrast dyes used in CT imaging. The cross-modality image synthesis from MRI to CT offers a promising solution to the limitations inherent to CT. Cycle consistent adversarial networks(CycleGAN), a high-performing architecture in the unpaired image-to-image synthesis, designed to produce convincing synthetic CT images from MRI images of the same subject without the need for an actual CT scan. However, CycleGAN often struggles to learn richer or more abundant image features and is difficult to capture structural information.

Methods

To address these issues, we developed a novel capsule cycle generative adversarial network (Capsule-CycleGAN). Capsule-CycleGAN is specifically designed to handle scenarios where paired data is not available. We integrate the capsule architecture into the CycleGAN discriminator, departing from the conventional convolutional neural network approach.

Result

To boost the capability of capturing relevant features, we refined the entire objective function by integrating margin loss, identity loss, cycle consistency, and adversarial loss, and we have achieved significant improvements in quantitative evaluation measures.

Conclusion

The effectiveness of our method was validated through comprehensive evaluations, demonstrating superior performance compared to other leading approaches in the synthesizing of high-quality CT images from MRI input.