Deep Learning Reconstruction for Rapid Multi-Component Relaxometry

项目摘要 弛豫法是用于量化组织特性的最常用的MRI技术之一。多组分 弛豫测定法测量多个水组分的弛豫特性，从而提供敏感的 和用于评价组织（例如软骨）的组成和微观结构的特异性MR生物标志物， 髓磷脂然而，由于需要拟合复杂的噪声敏感的MR信号模型，因此多分量MR信号模型不适用于复杂的噪声敏感的MR信号模型。 弛豫标测需要长扫描时间的重复扫描，限制了其广泛的临床应用。目标 这项研究提案的目的是通过利用最新的深度学习技术开发一种新方法， 以快速、临床可行的采集实现准确和高质量的多分量弛豫标测。 虽然最近的许多深度学习重建研究都集中在静态MR图像的快速成像上， 尽管取得了可喜的成果，但深度学习在加速松弛映射方面的应用却受到了限制。在 在这个项目中，我们建议开发、优化和评估一种新的深度学习技术， 具有多组分弛豫特性的组织的准确表征和量化。基础上 作为我们新开发的快速成像深度学习方法的基础，我们提出的方法将 利用高效的端到端卷积神经网络将欠采样MR图像直接转换为 多分量松弛的精确参数映射。一种新的基于数值布洛赫模拟的算法 用于精确模拟多组分弛豫行为，以确保准确性、可靠性和 深度学习训练过程中的鲁棒性。生成对抗网络将被纳入，以进一步 增强重建性能，以确保高质量的多分量弛豫映射 加速率。该提案还将探索新的数据增强方法， 图像数据集来创建广泛推广的深度学习模型。这确保了提议的深度 学习方法可以应用于不同的放松类型（例如，T2、T1和T1ρ），甚至 如果有限的训练数据集可用。我们的建议包括两个具体目标：（i）开发基于模型的 用于快速多组分弛豫测定的深度学习方法，以及（ii）研究合成图像的使用 用于训练深度学习模型的数据集。总的假设是，所提出的重建技术 可以提供一个独特的机会，探索加速多组分弛豫，利用 最新的深度学习技术，产生了一个准确，高效和可靠的模型，可以广泛应用于 可概括的该项目的成功完成将提供一个临床适用的多组分 松弛技术，用于更好地研究，理解和分期疾病，如骨关节炎和 多发性硬化这一概念可以显着推进定量MRI的临床翻译。

PROJECT SUMMARY Relaxometry is among the most used MRI technique for quantifying tissue properties. Multi-component relaxometry measures the relaxation characteristics of multiple water components, thus delivers both sensitive and specific MR biomarkers for evaluating composition and microstructure of tissues such as cartilage and myelin. However, due to the need to fit a complicated noise-sensitive MR signal model, multi-component relaxation mapping requires repeated scans with a long scan time, limiting its widespread clinical use. The goal of this research proposal is to develop a novel method via leveraging the latest deep learning techniques for realizing accurate and high-quality multi-component relaxation mapping at a rapid, clinical feasible acquisition. While many recent deep learning reconstruction studies have focused on rapid imaging for static MR images with promising results, applications of deep learning for accelerated relaxation mapping have been limited. In this project, we propose to develop, optimize, and evaluate a new deep learning technique that enables accurate characterization and quantification of tissues with multi-component relaxation properties. Building on the foundation of our newly developed deep learning method for rapid imaging, our proposed approach will utilize an efficient end-to-end convolutional neural network to directly convert undersampled MR images into accurate parametric maps for multi-component relaxation. A novel numerical Bloch-simulation based algorithm is applied to precisely model the multi-component relaxation behavior to ensure accuracy, reliability, and robustness in the deep learning training process. Generative adversarial network will be incorporated to further enhance the reconstruction performance to ensure high-quality multi-component relaxation mapping at high acceleration rates. This proposal will also explore new data augmentation approaches by using synthetic image datasets to create a widely generalizable deep learning model. This ensures that the proposed deep learning method can be applied to different relaxation types (e.g., T2, T1 and T1ρ) in many body regions, even if limited training datasets are available. Our proposal includes two specific aims: (i) to develop model-based deep learning method for rapid multi-component relaxometry, and (ii) to investigate the use of synthetic image datasets for training deep learning model. The overall hypothesis is that the proposed reconstruction technique can offer a unique opportunity to explore the acceleration of multi-component relaxometry by leveraging the latest deep learning techniques, resulting in an accurate, efficient, and reliable model that can be widely generalizable. Successful completion of the project will provide a clinically applicable multi-component relaxometry technique for better studying, understanding, and staging diseases such as osteoarthritis and multiple sclerosis. This concept could significantly advance quantitative MRI for clinical translation.