Scholarlink

Abstract <p>Cardiac 3D cine MRI is constrained by long acquisition times and challenging motion-handling requirements. As a result, certain patient conditions, such as cardiac arrhythmias, prevent the use of 3D sequences for whole-heart assessments. For these patients, 2D real-time sequences offer an alternative imaging solution, as they enable artifact-free acquisitions of individual slices with good in-plane contrast and temporal resolution. To derive whole-heart reconstructions from multi-planar 2D real-time MRI, we propose CiNeVol, a dynamic implicit neural representation network adapted from the static NeSVoR model. CiNeVol takes ECG- and respiration belt-derived motion states as inputs and provides decoupled estimates of cardiac and respiratory motion, yielding spatiotemporally consistent 3D reconstructions at arbitrary motion states. As lightweight in architecture, the model allows for fast instance optimization without requiring high-resolution training data. Evaluated on a simulated MRI dataset, CiNeVol demonstrates competitive image quality for both free-breathing (SSIM=0.92, PSNR=28.83, NMSE=0.03) and breath-hold (SSIM=0.89, PSNR=27.81, NMSE=0.04) 3D cine reconstructions, with faster runtimes than NeSVoR (SSIM=0.88, PSNR=26.13, NMSE=0.05) and SVRTK (SSIM=0.86, PSNR=26.28, NMSE=0.05). Preliminary in-vivo evaluations on three healthy volunteers confirm good spatiotemporal coherence, as reflected by superior expert scores, although remaining limitations, such as underestimation of cardiac contraction and image blurring, suggest the need for further optimization.</p> Graphical abstract <p>This work presents an implicit neural representation network that derives motion-corrected whole-heart reconstructions from multi-planar 2D real-time MRI. The model was evaluated on simulated cine MRI and in-vivo MRI acquisitions from three healthy volunteers. It takes ECG- and respiration belt-derived motion states as inputs and provides decoupled estimates of cardiac and respiratory motion, yielding spatiotemporally consistent 3D reconstructions at arbitrary motion states. By leveraging instance optimization, the model eliminates the need for high-resolution training data, while its lightweight architecture ensures competitive computation times.</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

Cardiac neural representations for ECG-guided slice-to-volume reconstruction

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • Nora Vogt,
  • Karyna Isaieva,
  • Jean-Sébastien Louis,
  • Pierre-André Vuissoz,
  • Jacques Felblinger,
  • Julien Oster

摘要

Abstract

Cardiac 3D cine MRI is constrained by long acquisition times and challenging motion-handling requirements. As a result, certain patient conditions, such as cardiac arrhythmias, prevent the use of 3D sequences for whole-heart assessments. For these patients, 2D real-time sequences offer an alternative imaging solution, as they enable artifact-free acquisitions of individual slices with good in-plane contrast and temporal resolution. To derive whole-heart reconstructions from multi-planar 2D real-time MRI, we propose CiNeVol, a dynamic implicit neural representation network adapted from the static NeSVoR model. CiNeVol takes ECG- and respiration belt-derived motion states as inputs and provides decoupled estimates of cardiac and respiratory motion, yielding spatiotemporally consistent 3D reconstructions at arbitrary motion states. As lightweight in architecture, the model allows for fast instance optimization without requiring high-resolution training data. Evaluated on a simulated MRI dataset, CiNeVol demonstrates competitive image quality for both free-breathing (SSIM=0.92, PSNR=28.83, NMSE=0.03) and breath-hold (SSIM=0.89, PSNR=27.81, NMSE=0.04) 3D cine reconstructions, with faster runtimes than NeSVoR (SSIM=0.88, PSNR=26.13, NMSE=0.05) and SVRTK (SSIM=0.86, PSNR=26.28, NMSE=0.05). Preliminary in-vivo evaluations on three healthy volunteers confirm good spatiotemporal coherence, as reflected by superior expert scores, although remaining limitations, such as underestimation of cardiac contraction and image blurring, suggest the need for further optimization.

Graphical abstract

This work presents an implicit neural representation network that derives motion-corrected whole-heart reconstructions from multi-planar 2D real-time MRI. The model was evaluated on simulated cine MRI and in-vivo MRI acquisitions from three healthy volunteers. It takes ECG- and respiration belt-derived motion states as inputs and provides decoupled estimates of cardiac and respiratory motion, yielding spatiotemporally consistent 3D reconstructions at arbitrary motion states. By leveraging instance optimization, the model eliminates the need for high-resolution training data, while its lightweight architecture ensures competitive computation times.