Ultra-Fast Knee MRI with Deep Learning

摘要 快速、稳健且可靠的定量膝关节 MR 成像将是研究关节的重要一步 退化、损伤和骨关节炎（OA）。 Automation of compositional and morphological feature extraction 这是将有希望的定量研究转化为临床实践的重要一步。 技术。 It would enable the analysis of large patient cohorts and assist the radiologist/clinician in augmenting MRI 的价值。 Automation of several human tasks has been achieved in the last few years by the usage of Deep Learning 技术。 With the availability of large amounts of annotated data and processing power, using the concepts 通过观察例子将数据转化为知识，监督学习今天可以完成 前所未有的挑战。除了图像分析和解释之外，深度学习还 彻底改变管道的获取和重建方面。模型可以学习直接映射 欠采样 k 空间和图像域之间。 While Deep Learning application to musculoskeletal imaging showed promising results when applied in a controlled setting, it is well understood that generalization beyond the statistical distribution of the training set is 仍然是一个未解决的挑战。在 MRI 中，当对经过训练的模型进行测试时，这会导致性能不佳 不同的成像协议或在不同的 MRI 系统上采集的图像。 通过这项提案，我们的目标是利用最近的进展并填补现有的空白。我们的目标是学习 新颖的集成模型能够同时加速 MRI 采集和自动化图像处理 that can overcome the limitation of single domain application.快速图像采集和准确的图像发布 processing are typically considered to be separate problems.然而，神经网络优化设计 使我们有机会将两者整合起来，以最大限度地提高加速和基于机器的图像处理 和解释。 We will use both publicly available benchmark dataset (FastMRI) and internally collected dataset to build deep learning models able to accurately reconstruct under sampled MRI acquisitions.我们会 use a dataset prospectively acquired during the course of this study to validate the clinical applicability of the 开发的方法。 Specifically, we will test the hypothesis that the proposed integrated pipeline can be applied in clinical setting for a fast and intelligent knee scan obtaining image quality comparable to standard acquisition and automated processing accuracy comparable with human reproducibility. 此外，我们建议将带注释的图像数据集和训练模型作为共享资源， centralized, open evaluation platform for MRI reconstruction and image post processing techniques.

ABSTRACT Fast, robust and reliable quantitative knee joint MR imaging would be a significant step forward in studying joint degeneration, injury and osteoarthritis (OA). Automation of compositional and morphological feature extraction of the tissues in the knee it is an essential step for translation to clinical practice of promising quantitative techniques. It would enable the analysis of large patient cohorts and assist the radiologist/clinician in augmenting the value of MRI. Automation of several human tasks has been achieved in the last few years by the usage of Deep Learning techniques. With the availability of large amounts of annotated data and processing power, using the concepts of transforming data to knowledge by the observation of examples, supervised learning can today accomplish challenges never demonstrated before. In addition to image analysis and interpretation, Deep Learning is revolutionizing the acquisition and reconstruction aspects of the pipeline. Models can learn a direct mapping between under sampled k-space and image domain. While Deep Learning application to musculoskeletal imaging showed promising results when applied in a controlled setting, it is well understood that generalization beyond the statistical distribution of the training set is still an unmet challenge. In MRI this translates into poor performances when trained models are tested on different imaging protocols or images acquired on different MRI systems. With this proposal, we aim to leverage on this recent advancement and filling the existing gaps. We aim to study novel integrated models able to simultaneously accelerate MRI acquisition and automate the image processing that can overcome the limitation of single domain application. Fast image acquisition and accurate image post processing are typically considered to be separate problems. However, the neural networks optimization design gives us an opportunity to integrate the two to maximize both acceleration and machine-based image processing and interpretation. We will use both publicly available benchmark dataset (FastMRI) and internally collected dataset to build deep learning models able to accurately reconstruct under sampled MRI acquisitions. We will use a dataset prospectively acquired during the course of this study to validate the clinical applicability of the developed methods. Specifically, we will test the hypothesis that the proposed integrated pipeline can be applied in clinical setting for a fast and intelligent knee scan obtaining image quality comparable to standard acquisition and automated processing accuracy comparable with human reproducibility. Additionally, we propose to make our annotated image datasets and trained models a shared resource, a centralized, open evaluation platform for MRI reconstruction and image post processing techniques.