Image analysis and machine learning methods for advanced MRI-neuropathology characterization

图像分析和机器学习方法代表了一种范式转变，在许多领域发挥着越来越重要的作用。这些方法在医学成像中的潜力包括疾病预测、个性化和病理分级。该研究计划的重点是推进病理分类技术。具体而言，我们的目标是开发新的图像纹理分析和机器学习方法，用于准确检测脑白色病变的损伤和修复，沿着数字病理学的验证。目前，我们的目标应用是磁共振成像（MRI）;然而，我们的方法可以应用于其他多个科学和工程领域，如地球科学，物理学，工程学和生物学。在当前的发现资助（2012-2018）下，我们开发了使用局部空间频率分析来理解MRI纹理的多分辨率模式的方法和工具。我们已经表明，一种特殊类型的白色物质结构与一系列独特的纹理谱有关，反映了该结构独特的形成尺度和排列方向。此外，MRI纹理的异质性与病理学的严重程度密切相关，如使用多发性硬化症作为示例所示。然而，在个体病变病理学的建模和验证方面仍然存在关键的未满足的需求。** 目标 * 本DG的总体目标是开发新的纹理分析和机器学习方法，以准确有效地映射神经病理学，特别是不同类型脑白色病变的损伤和修复程度。这包括5个目标：* 目标1：开发提取最关键的MRI纹理谱的新方法 * 目标2：发现在MRI中最佳检测组织方向性的方法 * 目标3：设计一种将纹理分析与机器学习相结合的方法 * 目标4：确定一种有效量化数字病理学的算法 * 目标5：建立一种融合MRI与数字病理学的方法，用于单一病变识别 * 方法 * 在目标1中，我们将研究新的光谱评估方法，并使用主成分分析识别最有用的纹理光谱。在目标2中，我们将评估相位一致性和其他与病变严重程度最相关的对齐评估方法。目标3将确定一种机器学习方法，用于基于新的纹理特征进行准确的病变类型分类。在目标4中，我们将开发新的图像分析和机器学习方法来量化病理图像，在目标5中，我们将比较MRI和病理测量，以建立一种评估单个病变属性的非侵入性方法。影响 * 新的纹理分析和机器学习方法在成像中的应用将导致信号处理、地球科学、组织工程和医疗保健等各个领域的先进技术的发展。

Introduction***Image analysis and machine learning methods represent a paradigm shift and are playing an increasingly critical role in many fields. The potential of these methods in medical imaging includes disease prediction, personalization, and pathological grading. This research program focuses on advancing technologies to categorize pathology. Specifically, we aim to develop new image texture analysis and machine learning methods for accurate detection of injury and repair in brain white matter lesions, along with validation by digital pathology. Currently our target applications are in magnetic resonance imaging (MRI); however, our approaches can be applied in multiple other science and engineering fields such as geoscience, physics, engineering, and biology.******Under the current Discovery Grant (2012-2018), we have developed methods and tools to understand the multi-resolution pattern of MRI texture using localized spatial frequency analysis. We have shown that a specific type of white matter structure is related to a unique series of texture spectra, reflecting the unique forming scales and aligning directions of the structure. Further, the heterogeneity of MRI texture correlates strongly with the severity of pathology as shown using multiple sclerosis as an example. However, there are still critical unmet needs in both modeling and verification of individual lesion pathology. ******Objective***The overall objective of this DG is to develop new texture analysis and machine learning methods for accurate and efficient mapping of neuropathology, particularly the degree of injury and repair in different types of brain white matter lesions. This includes 5 aims:***Aim 1: Develop new methods to extract the most critical MRI texture spectra ***Aim 2: Discover approaches for best detecting tissue directionality in MRI***Aim 3: Design a method to integrate texture analysis with machine learning ***Aim 4: Identify an algorithm for efficient quantification of digital pathology ***Aim 5: Build a method to fuse MRI with digital pathology for single lesion identity******Approach***In Aim 1, we will investigate new spectrum assessing methods and identify the most useful texture spectra using principal component analysis. In Aim 2, we will evaluate phase congruency and other alignment assessing methods that best correlate with lesion severity. Aim 3 will determine a machine learning method for accurate lesion type classification based on new texture features. In Aim 4, we will develop new image analysis and machine learning methods for quantifying pathology images, and in Aim 5 we will compare MRI and pathological measures to establish a non-invasive method for assessing single lesion property.******Impact***The application of new texture analysis and machine learning methods to imaging will lead to the development of advanced technologies for various fields such as signal processing, geoscience, tissue engineer, and healthcare.