Deep Learning-based Imaging Biomarkers for Knee Osteoarthritis

摘要 在美国，60多万例与膝骨性关节炎(OA)相关的全膝关节置换(TKR)病例是 每年报告，估计每年直接成本超过170亿美元。人们对疾病的需求越来越大-- 修改治疗方法，防止或推迟TKR的需要。然而，这种疗法的发展仍然 具有挑战性，因为缺乏客观和可测量的骨关节炎生物标记物来预测疾病的进展。这一过程 个人之间的办公自动化差异很大，而且办公自动化进展太慢，很难识别 敏感的骨性关节炎生物标志物能够捕捉膝关节的微小变化。这减缓了……的发展。 有效的治疗方法，并防止医生提供最有效的建议，将需求降至最低 为了TKR。在这个项目中，我们的目标是开发成像生物标记物来监测膝盖骨关节炎相关的轻微变化。 导致TKR的关节健康。为了实现这一目标，我们将把新的深度学习算法与临床相结合 和来自骨关节炎倡议(OAI)的成像数据。OAI数据集包括临床数据、生物标本、 8年来收集的X光照片和磁共振(MR)图像。拟议的项目有三个 具体目标：(1)从双边开发一个自动化的与OA相关的生物标记物识别工具 深层卷积神经网络用于膝关节后前固定屈曲X线片 神经网络(RNN)与受试者(n=882)的OA进展结果相结合；(Ii)开发 使用3D从结构和成分MR图像中自动识别与OA相关的生物标志物的工具 CNN和RNN结合受试者的OA进展结果(n=882)；以及(Iii)确定 基于深度学习的成像生物标记物可以作为使用受试者预测OA进展的替代物 队列(n=296)，独立于用于识别成像生物标记物的队列。拟议中的项目将与 使用诊断放射学进行深度学习，直接从图像中揭示与OA相关的关键功能组合 只需最少的用户交互。这将有助于使用整体评估快速个性化评估办公自动化进展 直接获取膝关节图像。如果成功，这项研究将为成像技术的发展带来新的见解 生物标记物的进展，并更广泛地进入我们的理解和治疗的OA。《知识》 在这个项目中获得的成果将有助于通过打开新的视角来推进对办公自动化进展的密切监测 干预研究和临床实践中潜在纳入的区域和参数。

ABSTRACT In the U.S., more than 600,000 knee osteoarthritis (OA)-related total knee joint replacement (TKR) cases are reported every year, exceeding $17 billion estimated direct costs annually. There is a growing need for disease- modifying therapies that prevent or delay the need for TKR. However, development of such therapies remains challenging due to the lack of objective and measurable OA biomarkers for disease progression. The course of the OA is highly variable between individuals and the OA progresses too slowly, making it difficult to identify sensitive OA biomarkers capable of capturing minor changes on the knee joint. This has slowed development of effective therapies and prevents physicians from providing the most effective advice about minimizing the need for TKR. In this project, our goal is to develop imaging biomarkers to monitor minor OA-related changes in knee joint health that lead to TKR. To achieve this goal, we will combine novel deep learning algorithms with clinical and imaging data from the Osteoarthritis Initiative (OAI). The OAI dataset includes clinical data, biospecimens, radiographs, and magnetic resonance (MR) images collected over 8 years. The proposed project has three Specific Aims: (i) to develop an automated OA-relevant biomarker identification tool from the bilateral posteroanterior fixed-flexion knee radiographs using deep convolutional neural networks (CNNs) and recurrent neural networks (RNNs) combined with the OA progression outcome of subjects (n = 882); (ii) to develop an automated OA-relevant biomarker identification tool from structural and compositional MR images using 3D CNNs with RNNs combined with the OA progression outcome of subjects (n = 882); and (iii) to determine whether deep learning–based imaging biomarkers can act as surrogates to predict the OA progression using a subject cohort (n = 296) independent of the cohort used to identify imaging biomarkers. The proposed project will couple deep learning with diagnostic radiology to unveil key combinations of OA-relevant features directly from images with minimal user interaction. This will facilitate fast individualized assessment of OA progression using whole knee joint images directly. If successful, this study will bring new insights into the development of imaging biomarkers for OA progression and more broadly into our understanding and treatment of OA. The knowledge gained in this project will help to advance close monitoring of OA progression by opening new perspectives on the regions and parameters for potential inclusion in both intervention studies and clinical practice.