A comprehensive deep learning framework for MRI reconstruction

项目摘要/摘要 本调查的主要目标是开发和验证全面、稳健的深度学习(DL) 改进磁共振成像重建的框架，超越现有技术的限制。拟议的框架 使用即插即用算法将物理驱动的MR采集模型与最先进的学习相结合 图像模型，由图像去噪子程序实例化。要充分利用MR的丰富结构 对于图像，我们建议使用基于DL的去噪器，这些去噪器是以应用程序特定的fic方式训练的。建议数 框架，称为PNP-DL，提供了比其他现有的DL方法以及压缩感知更好的优点 (CS)。与现有的用于MRI重建的DL方法相比，PNP-DL对不可避免的变化具有更强的免疫力 在正向模型中，如线圈灵敏度或欠采样模式的变化，允许其泛化 跨应用程序和采集设置。与CS相比，PNP-DL恢复图像更快，质量更高， 具有潜在的更高的诊断价值。 我们的初步结果突出了PNP-DL在推进MRI技术方面的潜力。在这项工作中，我们将- 进一步开发PNP-DL，并在以下主要应用中进行验证：心脏电影、2D脑和3D脑成像。 在目标1中，我们将训练和优化基于卷积神经网络的应用程序特定的fic++去噪器。 上述申请。选择去噪性能最好的去噪器进行进一步的研究 调查。在目标2中，我们将开发并比较不同的PNP算法。产生最佳结果的算法 重建精度和计算速度的结合将在Gadgetron中实现，用于内联 正在处理。在目标3中，我们将比较PNP-DL与其他使用Retro-Dl的最先进方法的性能。 特别是抽样不足的数据。这项研究将证明，在图像质量方面，PNP-DL优于 CS和现有的DL方法，尽管有更高的加速，但并不逊色于具有Rate-2加速器的并行MRI。 阿齐兹。在目标4中，我们将使用预期欠采样的成人数据来评估PNP-DL的性能 和儿科病人。该项目的成功完成将证明PNP-DL的性能优于STATE- 在图像质量方面的最先进方法，同时表现出一定程度的稳健性和广泛的适用性， 避开了其他基于DL的MRI重建方法。加速和图像质量的提高 这些发展将使fi几乎所有的核磁共振应用受益，包括儿科成像，在 减少镇静是一种迫切的需求，而高维成像应用(例如，全心4Dfl 成像)，对于常规的临床使用来说太慢了。

PROJECT SUMMARY/ABSTRACT The primary goal of this investigation is to develop and validate a comprehensive, robust deep learning (DL) framework that improves MRI reconstruction beyond the limits of existing technology. The proposed framework uses “plug-and-play” algorithms to combine physics-driven MR acquisition models with state-of-the-art learned image models, which are instantiated by image denoising subroutines. To fully exploit the rich structure of MR images, we propose to use DL-based denoisers that are trained in an application-specific manner. The proposed framework, termed PnP-DL, offers advantages over other existing DL methods, as well as compressed sensing (CS). Compared to existing DL methods for MRI reconstruction, PnP-DL is more immune to inevitable variations in the forward model, such as changes in the coil sensitivities or undersampling pattern, allowing it to generalize across applications and acquisition settings. Compared to CS, PnP-DL recovers images faster, with higher quality, and with potentially superior diagnostic value. Our preliminary results highlight the potential of PnP-DL to advance MRI technology. In this work, we will fur- ther develop PnP-DL and validate it in these major applications: cardiac cine, 2D brain, and 3D brain imaging. In Aim 1, we will train and optimize convolutional neural network-based application-specific denoisers for the above-mentioned applications. The denoiser with the best denoising performance will be selected for further investigation. In Aim 2, we will develop and compare different PnP algorithms. The algorithm yielding the best combination of reconstruction accuracy and computational speed will be implemented in Gadgetron for inline processing. In Aim 3, we will compare the performance of PnP-DL to other state-of-the-art methods using retro- spectively undersampled data. This study will demonstrate that, in terms of image quality, PnP-DL is superior to CS and existing DL methods and, despite higher acceleration, is non-inferior to parallel MRI with rate-2 acceler- ation. In Aim 4, we will evaluate the performance of PnP-DL using prospectively undersampled data from adult and pediatric patients. Successful completion of this project will demonstrate that PnP-DL outperforms state- of-the-art methods in terms of image quality while exhibiting a level of robustness and broad applicability that has eluded other DL-based MRI reconstruction methods. The acceleration and image quality improvement afforded by these developments will benefit almost all MRI applications, including pediatric imaging, where reducing sedation is a pressing need, and high-dimensional imaging applications (e.g., whole-heart 4D flow imaging), which are too slow for routine clinical use.