A Bayesian Framework for Inter-Modality Deformable Image Registration Evaluation

描述（由申请人提供）：随着医学影像文献中不断报道新的创新和日益复杂的图像处理技术，在关键评估和质量控制方法方面缺乏同步的复杂性。尽管有许多关于新的DIR算法及其潜在的诊断和治疗医学应用的报道，但科学文献缺乏针对医学应用的DIR性能评估、比较测试和验证的标准化程序。专家确定的解剖特征对有可能成为评估DIR空间精度的广泛采用的参考；然而，它们的使用仍然有很大的可变性。分析匹配地标对的统计方法仅限于描述性统计，总结了测量的配准误差，未能考虑解剖定位的不确定性，观察者之间的可变性以及图像空间的体素离散化。贝叶斯方法在解释现代医学研究数据中的效用早已得到认可。就我们的目的而言，贝叶斯方法的优势在于它允许从多个来源得出关于算法性能特征的判断，包括用于特征对定位的多个观察者、多种成像模式和独立参考数据集。这有助于解释测量数据，并使我们能够将成像获取和重建过程的知识纳入反映潜在物理过程的先验分布的公式中。这使得算法的空间精度表现比现在更完整。提出的研究目标是开发一个计算框架和软件基础设施，用于变形图像配准空间精度的贝叶斯分析。执行这些分析的软件将被整合到一个公开可用的参考图像数据库中，允许研究人员在目前缺乏的标准分析框架内定量评估和比较公共数据集上的多种图像配准算法/实现。本课题的具体研究目的是：1。创建一个案例参考库，以测量DIR空间精度性能和模态间（CT-MRI）配准的不确定性。2. 利用专家选择的地标特征方法开发并验证了用于DIR空间精度评估的贝叶斯层次模型。3. 推广标准化的DIR空间精度贝叶斯分析软件。广泛适用的DIR评价通用数据集的可用性将促进科学文献的精简比较评价和荟萃分析，并为开发目前缺乏的标准化评价方法提供基础。此外，人们对采用多模态方法进行治疗前放疗（RT）计划和图像引导的RT传递非常感兴趣，其中磁共振成像（MRI）的优越采集和软组织特征与计算机断层扫描（CT）固有的电子密度信息和几何保真度相结合。纳入CT-MRI参考数据将允许研究者探索多模式RT计划和图像引导交付方法的可行性，这需要对互补数据集进行精确的空间配准。通过提供一个严格的计算框架，将解剖特征对用于DIR评估的不确定性纳入其中，本研究有可能为医学成像中DIR的临床验证、验收测试和质量保证制定未来的方案指南。

DESCRIPTION (provided by applicant): As new innovative and increasingly sophisticated image processing techniques are continually reported in the medical imaging literature, concurrent sophistication in methods for critical evaluation and quality control is lacking. Despit numerous reports of novel DIR algorithms and their potential diagnostic and therapeutic medical applications, the scientific literature is lacking standardized procedures for DIR performance evaluation, comparison testing, and validation specific to medical application. Expert-determined anatomic feature-pairs have the potential to become a widely adopted reference for evaluating DIR spatial accuracy; however, there is still great variability in their use. Statistical methods fr analyzing the matched landmark pairs have been limited to descriptive statistics summarizing the measured registration errors, failing to account for uncertainty in anatomic localization, variability among observers, and voxel discretization of the image space. The utility of Bayesian methods in the interpretation of modern medical research data has long been recognized. For our purposes, the strength of a Bayesian approach is one that allows judgment regarding an algorithm's performance characteristics to be derived from multiple sources, including multiple observers for feature-pair localization, multiple imaging modalities, and independent reference datasets. This facilitates interpretation of the measured data, and allows us to incorporate knowledge of the imaging acquisition and reconstruction process into formulation of prior distributions reflective of the underlying physical processes. This results in a more complete representation of an algorithm's spatial accuracy performance than is available today. The goal of the proposed research is to develop a computational framework and software infrastructure for Bayesian analysis of deformable image registration spatial accuracy. Software for performing these analyses will be incorporated into a publicly available reference image database, allowing investigators to quantitatively evaluate and compare multiple image registration algorithms/implementations on a common dataset, within a standard analysis framework that is currently lacking. The Specific Aims of the proposed research are: 1. Create a reference library of cases to measure DIR spatial accuracy performance and uncertainty for inter-modality (CT-MRI) registration. 2. Develop and validate a Bayesian hierarchical model for DIR spatial accuracy evaluation using the expert selected landmark feature approach. 3. Disseminate software for standardized Bayesian analysis of DIR spatial accuracy. The availability of a common dataset for DIR evaluation that is broadly applicable will facilitate streamlined comparative evaluation and meta-analysis of the scientific literature, and provide a foundation upon which to develop a standardized evaluation methodology that is presently lacking. Additionally, there is much interest to adopt a multi-modality approach to pre-treatment radiotherapy (RT) planning and image guided RT delivery, in which the superior acquisition and soft-tissue characteristics of magnetic resonance imaging (MRI) are integrated with the electron density information and geometric fidelity inherent to computed tomography (CT). Inclusion of CT-MRI reference data will allow investigators to explore feasibility of a multi-modal approach to RT planning and image-guided delivery, which requires accurate spatial registration of the complementary datasets. By providing a rigorous computational framework for incorporating uncertainty in the use of anatomic feature-pairs for DIR evaluation, the proposed study has the potential to shape future protocol guidelines for clinical validation, acceptance testing, and quality assurance of DIR in medical imaging.