Develop a large-scale library of comprehensive deformable image registration (DIR) benchmark datasets and an integrated framework for quantifying accuracy of patient-specific DIR results

总结 从同一患者采集的不同图像集之间的可变形图像配准（Deformable Image Registration，简称RDR）是一种图像配准方法。 对于许多重要的诊断和治疗任务，例如肿瘤诊断， 评价肿瘤对治疗的反应和图像引导手术。可编程算法计算组织 通过最大化运动图像和目标图像之间的强度和/或结构相似性来变形，以及 变形规律精确度，这是体素级的位置之间的对应关系 两个图像，是不能保证的，往往是不充分的，不可预测的和病人的具体。精确度是 这在很大程度上取决于解剖部位、成像模态和质量、算法设计 实施、操作员技能和工作流程选择。不准确的DIR可能会导致严重的 有害影响临床决策、治疗质量和患者安全。对当前的信心不足 配准工具极大地限制了在自动化临床决策中更广泛地使用RFID 任务和改善诊断和治疗结果。 我们认为，缺乏准确或稳健的性能是由于当前的DIR算法 是基于组织变形和失败的过于简单的模型，以适应 CT图像质量。目前，不存在用于定量地和自动地评估患者- 具体的精确度。因此，我们有动力进行两项研究： 1)构建一个大型而全面的BLOG基准数据集库，以支持BLOG算法 在具有挑战性的环境中进行验证。每一个新的数据集将包括自动和 精确检测的界标对、小血管截面对和器官分割， 大血管目前还没有这样的数据集。这些数据集将刺激 开发能够支持复杂的患者特定组织的新的和先进的并行计算算法 变形这些数据集将对超越物联网的应用程序非常有价值，例如 语义分割和血管提取等。 2)开发用于定量验证患者特定DIR的综合方法。该自动化设备 验证程序将使用多个新的深度学习模型，用于自动器官 分割、血管分叉检测和TREs（目标）的3D矢量场的直接预测 注册错误）。这些待开发的基于深度学习的图像处理程序是 对于图像噪声和强度变化是鲁棒的，并且将自然地支持许多 解剖部位。该验证程序将为患者特定的 用于支持临床应用的DIR。

Summary Deformable image registration (DIR) between different image sets acquired from the same patient is a key enabling technology for many important diagnostic and therapeutic tasks, e.g. tumor diagnosis, evaluation of tumor response to treatment, and image-guided surgery. DIR algorithms compute tissue deformation by maximizing intensity and/or structural similarity between moving and target images, and regularity of deformation. DIR accuracy, which is the voxel-level positional correspondence between the two images, is not guaranteed, often inadequate, unpredictable and patient specific. DIR accuracy is largely dependent on anatomical site, image modality and quality, algorithm designs and implementations, operator skills and workflow selections. Inaccurate DIRs can have significant deleterious impact clinical decisions, treatment quality and patient safety. Lack of confidence in current registration tools has significantly limited the broader use of DIR in automating clinical decision-making tasks and improving diagnostic and therapeutic outcomes. We posit that lack of accurate or robust performance arises from the fact that current DIR algorithms are based upon overly simplistic models of tissue deformation and failure to accommodate the reality of CT image quality. Currently, no method exists for quantitatively and automatically evaluating patient- specific DIR accuracy. We are therefore motivated to conduct two studies: 1) Build a large and comprehensive library of DIR benchmark datasets to support DIR algorithm validation in challenging settings. Each new DIR benchmark dataset will consist of automatically and precisely detected landmark pairs, small blood vessel section pairs, and segmentation of organs and large blood vessels. Currently no such DIR benchmark dataset exist. These datasets will spur development of new and advanced DIR algorithms able to support complex, patient-specific tissue deformation. These datasets will be immensely valuable for applications beyond DIR such as semantic segmentation and vessels extraction, etc. 2) Develop integrated methods for quantitative verification of patient-specific DIRs. The automatic DIR verification procedure will use multiple novel deep-learning models for automatic organ segmentation, vessel bifurcation detection and direct prediction of 3D vector field of TREs (target registration error). These to-be-developed deep-learning-based image processing procedures are robust with respect to image noise and intensity variations, and will naturally support many anatomical sites. This DIR verification procedure will provide quality assurance for patient-specific DIRs for supporting clinical applications.