Learning an Optimized Variational Network for Medical Image Reconstruction

项目摘要 我们提出了一种基于深度学习和计算机视觉的医学图像重建新方法， 模拟了人类放射科医生如何利用多年的经验，从阅读数千例病例中， 识别解剖结构、病理和图像伪影。我们的方法是基于一个新的想法， 变分网络，它在深度学习中嵌入了广义压缩感知概念， 框架.我们建议学习一个完整的重建过程，包括过滤器内核和惩罚 函数来区分真实图像内容和伪影，通常必须调整的所有参数 手动以及由该变分网络描述的相关数值算法。训练步骤 与时间关键的图像重建步骤解耦，然后可以近实时地执行 不中断临床工作流程。我们从加速磁共振得到的初步病人数据 成像（MRI）采集表明，我们的学习方法优于目前最先进的 并且相对于在日常临床中出现的变化是鲁棒的 成像情况。在我们的第一个目标中，我们将测试这样一个假设，即学习可以这样进行， 对于数据采集的变化具有鲁棒性。在第二个目标中，我们将回答这个问题，如果可能的话， 学习多种MR成像应用的单一重建程序。最后，我们将进行临床 对300名因膝关节内部紊乱而接受成像的患者进行的阅片者研究。我们将比较我们的 提出了一种临床标准重建方法。我们的假设是，我们的方法将导致 当使用5分钟检查时，相同的临床诊断和患者管理决策。 该项目将加速成像应用于具有广泛公共卫生影响的应用，从而改善 临床结果和降低医疗成本。此外，从发展中获得的见解， 该项目将回答新兴机器学习领域目前最重要的开放问题。 用于医学图像重建。最后，鉴于这一领域的活动最近有所增加， 对可用于培训的原始k空间数据的公开可用数据存储库的需求很大 和验证。由于本项目将获得的所有数据都将用于研究， 社区，这个项目将是满足这一需求的第一步。

Project Summary We propose a novel way of reconstructing medical images rooted in deep learning and computer vision that models the process how human radiologists are using years of experience from reading thousands of cases to recognize anatomical structures, pathologies and image artifacts. Our approach is based on the novel idea of a variational network, which embeds a generalized compressed sensing concept within a deep learning framework. We propose to learn a complete reconstruction procedure, including filter kernels and penalty functions to separate between true image content and artifacts, all parameters that normally have to be tuned manually as well as the associated numerical algorithm described by this variational network. The training step is decoupled from the time critical image reconstruction step, which can then be performed in near-real-time without interruption of clinical workflow. Our preliminary patient data from accelerated magnetic resonance imaging (MRI) acquisitions suggest that our learning approach outperforms the state-of-the-art of currently existing image reconstruction methods and is robust with respect to the variations that arise in a daily clinical imaging situation. In our first aim, we will test the hypothesis that learning can be performed such that it is robust against changes in data acquisition. In the second aim, we will answer the question if it is possible to learn a single reconstruction procedure for multiple MR imaging applications. Finally, we will perform a clinical reader study for 300 patients undergoing imaging for internal derangement of the knee. We will compare our proposed approach to a clinical standard reconstruction. Our hypothesis is that our approach will lead to the same clinical diagnosis and patient management decisions when using a 5min exam. The immediate benefit of the project is to bring accelerated imaging to an application with wide public-health impact, thereby improving clinical outcomes and reducing health-care costs. Additionally, the insights gained from the developments in this project will answer the currently most important open questions in the emerging field of machine learning for medical image reconstruction. Finally, given the recent increase of activities in this field, there is a significant demand for a publicly available data repository for raw k-space data that can be used for training and validation. Since all data that will be acquired in this project will be made available to the research community, this project will be a first step to meet this demand.