Large-scale automatic analysis of the OAI magnetic resonance image dataset

摘要 骨关节炎（OA）是关节炎最常见的形式，也是残疾的常见原因。虽然OA影响 仅在美国就有数百万人，关节置换通常是唯一可用的治疗方法， 疾病的痛苦和残疾变得太大。OA研究和临床护理的进展已经 由于缺乏敏感的生物标志物和缺乏用于检测这种生物标志物的分析方法， 生物标志物在一些现有的大型数据集，如骨关节炎倡议（OAI）的数据集。 OAI的磁共振图像（MRI）数据集包含极有价值的纵向图像数据 从超过4,000名受试者收集了8年的时间。虽然软骨损失被认为是 骨关节炎的主要因素，迄今为止，软骨分割公开提供的图像只有约1% OAI数据集。这严重限制了对膝关节软骨变化及其与预后关系的研究 措施获得完整数据集的基于图像的软骨生物标志物是困难的，因为大多数现有的分析 这些方法充其量是半自动化的。一个关键的挑战是，现有的方法不能大规模地扩展 数据集：无论是从财务上（这种分析将花费数百万美元），还是从实用的角度来看， 例如，在一个实施例中，人工分割软骨可能需要一个人十年的全职工作。 该项目有两个目标： 1)我们将发明先进的图像分析和统计方法， OAI MRI数据集的分析，即，将允许我们分析完整的OAI数据集。这些办法将 包括自动分割和表征膝关节软骨以及评估 主题和时间。我们所有的分析软件都将以开源的形式向公众开放， 免费为任何人使用。我们将支持定制计算集群、云计算和并行计算。 2)通过促进对整个数据集的大规模分析，所提出的方法将使我们能够重新审视 许多重要的临床问题由于现有方法中的差距而没有解决。特别是，标准的放射照相 OA进展的结局指标（基于Kelling-Lawrence分级和/或关节间隙狭窄） 可靠性低，难以解释，对变化的反应差。因此，我们将探索局部软骨 厚度作为OA进展的指标及其与OA假定风险因素的相关性， （与预期相反）仅显示出与放射学检查的有限、冲突或不确定的关联 措施我们还将研究从短期软骨损伤预测OA的长期进展。 这些特征可以帮助识别软骨快速丧失风险最高的个体。一旦确定， 然后这些个体可以成为更积极的治疗或临床试验的目标。

ABSTRACT Osteoarthritis (OA) is the most frequent form of arthritis and a common cause of disability. While OA affects millions of people in the United States alone, joint replacement is generally the only available treatment when the pain and disability of the disease become too great. Advances in OA research and clinical care have been greatly hindered by a lack of sensitive biomarkers and by the absence of analysis methods for detecting such biomarkers in some existing large datasets, such as the dataset of the Osteoarthritis Initiative (OAI). The magnetic resonance image (MRI) dataset of the OAI contains extremely valuable longitudinal image data from more than 4,000 subjects collected over an 8-year period. While cartilage loss is believed to be the dominating factor in OA, to date cartilage segmentations are publicly available for only about 1% of the images of the OAI dataset. This severely limits research on knee cartilage changes and their relation to outcome measures. Obtaining image-based cartilage biomarkers for the full dataset is difficult, as most existing analysis approaches are at best semi-automated. A key challenge is that the existing approaches do not scale to large datasets: neither financially (such analysis would cost millions of dollars) nor from a practical point of view – e.g., manually segmenting cartilage would likely require a decade of full-time work by one individual. The aim of this project is two-fold: 1) We will invent advanced image-analysis and statistical approaches which will allow for truly large-scale analysis of the OAI MRI dataset, i.e., will allow us to analyze the full OAI dataset. These approaches will include methods to automatically segment and characterize knee cartilage and to assess differences between subjects and across time. All our analysis software will be made available in open-source form to the public, free to use for anybody. We will support custom compute clusters, cloud- and parallel computing. 2) By facilitating large-scale analysis of the entire dataset, the proposed approaches will allow us to revisit many important clinical questions left open by gaps in prior methods. In particular, standard radiographic outcome measures for OA progression (based on Kellgren-Lawrence grade and/or joint space narrowing) have low reliability, are difficult to interpret, and respond poorly to change. We will therefore explore local cartilage thickness as a measure for OA progression and its associations with putative risk factors of OA, which (contrary to expectation) have only shown limited, conflicting, or inconclusive associations with radiographic measures. We will also investigate the prediction of long-term OA progression from short-term cartilage characteristics, which could help identify individuals at highest risk of rapid cartilage loss. Once identified, these individuals could then be targeted for more aggressive therapy or for clinical trials.