Deep Learning for Characterizing Knee Joint Degeneration Predicting Progression of Osteoarthritis and Total Knee Replacement

摘要 这项提案旨在开发深度学习方法，以自动提取形态成像 与膝关节骨关节炎（OA）和全膝关节置换术相关的特征。定量评价 磁共振成像（MRI）在临床环境中的OA研究中起着核心作用，MR报告通常 往往是主观的，定性的，流行病学研究中使用的分级方案没有使用 因为它们非常耗时，而且不适合当今的需求， 不断变化的医疗保健场景。我们面临的“大数据”挑战和机遇使我们有必要建立 使工具（i）能够自动提取形态学OA成像特征，目的是评估 以前从未探索过的更大样本量的疾病进展预测能力;（ii） 通过应用状态信息来发现未探索的数据驱动的成像特征， 现有技术的深度学习方法（1）;（iii）将多模态成像与临床、功能、活动相结合， 其他数据来定义OA中关节退变的轨迹。利用这些先进技术的力量 技术，并与注释图像的大型数据集的非凡可用性;在这个项目中，我们 建议开发一种能够分割肌肉骨骼组织的自动后处理管道， 在磁共振图像（MRI）中识别骨关节炎的形态学特征，如常用MRI所定义 评分系统关节中组织形态学分级的自动化将是一个重要的 OA研究和临床实践的突破。它将使分析大样本量， 辅助放射科医生/临床医生对图像进行分级，花费相对较短的时间，降低成本，以及 可以潜在地改进分类模型。自动管线的可用性， MRI中的形态学异常将彻底改变临床实践，包括半定量 等级，而不是放射学临床报告中的主观印象。在这项研究中，我们还旨在开发一种 完整的监督深度学习方法，以获得数据驱动的非线性和语义表示 基本特征之间的聚合能够利用隐藏在3D复杂性中的潜在信息， MR图像，消除了对所选特征的标称等级的需要。这第二个目标，虽然在高 风险也有潜在的异常高的影响;因为它偏离了经典的假设驱动的研究， 建立了一个新的平移平台，以彻底改变研究中MR图像的形态分级， 而且是范式转换，因为它可以为常规提供更定量的特征驱动基础， 放射临床报告。该提案的临床影响在于第三个目标（R33阶段），其中我们 建议将R61阶段开发的解决方案翻译为UCSF临床档案中的图像 （PACS），并计划在电子健康记录中包括人口统计和临床数据，以建立 定义全膝关节置换术的模型。

ABSTRACT This proposal aims to develop deep learning methods to automate the extraction of morphological imaging features relevant to knee osteoarthritis (OA), and total knee replacement. While, quantitative evaluation Magnetic Resonance Imaging (MRI) plays a central role in OA research in the clinical setting MR reports often tend to be subjective, qualitative, and the grading schemes utilized in epidemiological research are not used because they are extraordinarily time consuming and do not lend themselves to the demands of todays changing healthcare scenario. The “Big Data” challenge and opportunity facing us makes it necessary to build enabling tools (i) to automate the extraction of morphological OA imaging features, with the aim of evaluating disease progression prediction capabilities on larger sample sizes that have never been explored before; (ii) to discover latent patterns by uncovering unexplored data-driven imaging features by the application of state of the art deep learning approaches (1); (iii) combine multi-modality imaging with clinical, functional, activity, and other data to define the trajectory of joint degeneration in OA. Leveraging the power of these state of the art techniques, and with the extraordinary availability of a large datasets of annotated images; in this project, we propose to develop an automatic post-processing pipeline able to segment musculoskeletal tissues and identify morphological OA features in Magnetic Resonance Images (MRI), as defined by commonly used MRI grading systems. Automation of morphological grading of the tissues in the joint would be a significant breakthrough in both OA research and clinical practice. It would enable the analysis of large sample sizes, assist the radiologist/clinician in the grading of images, take a relatively short amount of time, reduce cost, and could potentially, improve classification models. The availability of automatic pipelines for the identification of morphological abnormities in MRI would drastically change clinical practice, and include semi-quantitative grades, rather than subjective impressions in radiology clinical reports. In this study, we also aim to develop a complete supervised deep learning approach to obtain data-driven representations as non-linear and semantic aggregation among elementary features able to exploit the latent information hidden in the complexity of a 3D MR images, eliminating the need for nominal grades of selected features. This second aim, while being at high risk has also a potential exceptional high impact; as it departs from the classical hypothesis driven studies, and builds a novel translational platform to revolutionize morphological grading of MR images in research studies, but also is paradigm-shifting in that it may provide a more quantitative feature driven basis for routine radiological clinical reports. The clinical impact of this proposal lies in the third aim (R33 phase), in which we propose to translate the solutions developed in the R61 phase on images in the UCSF clinical archives (PACS), and plan to include also demographic and clinical data in the electronic health records, to build the models defining total knee replacements.