Models and methods for automatically measuring disease body-wide and staging disease via FDG-PET/CT in Lymphoma

定量放射学极有希望改变我们诊断、监测、分期、预测、 并检测疾病以及计划和指导患者的治疗干预。然而，这一过程 医学图像中解剖器官和病理区域的定位和勾画，称为图像 自动化程度较高的细分仍然是这些进步的主要障碍。多数 图像分割的发展集中在特定的器官或一小群物体上 特定的身体区域。当任何一种新方法或现有方法的主要改编被设计时 这些参数发生了变化。这种方法是不可持续的，并成为在处理 对于全身系统疾病，需要全身图像分析。关键的进展是需要的 在这一领域要克服两个主要挑战：(1)尽管关于正常解剖的先前信息被认为是 对于图像分割和分析至关重要，它的创建和大规模应用在全国范围内还没有 已经尝试过了，但都严重缺乏。(2)使用这种信息的技术和方法 不存在高度自动化的疾病量化。首要目标是 就是要克服这些 开发一种全身可概括的解剖学引导的深度学习图像分割的挑战 方法论 并展示了其在弥漫性大B细胞淋巴瘤研究中的应用 (DLBCL)，基于PET的分期和反应评估是至关重要的。 该项目有三个具体目标。目的1：开发一个代表人体的全身解剖学模型家族 整个人类成年年龄段。 来自两个机构(宾夕法尼亚大学和宾夕法尼亚大学)的600名患者的现有FDG PET/CT扫描 纽约质子中心)覆盖10个年龄段，将被用来建立涉及50个器官的解剖模型 扩展躯干的50个淋巴结区，包括颈部、胸部、腹部和骨盆。 一家人 代表4个身体区域和10个年龄段的40个解剖模型将从大约60,000个3D 对象采样。AIM2：开发、实现和验证用于本地化对象的方法 量化疾病，而不明确描述器官和损伤。性别和年龄相关的解剖模型 将用于在任何给定的患者PET/CT图像中自动定位以上100个对象，并 量化每个身体区域、器官和淋巴结区的疾病。这些方法将在400PET/CT上进行测试 DLBCL患者的影像。目的：建立和验证DLBCL疾病自动分期方法 和预后。疾病数量信息将被用于开发自动分期和结果 预测算法将在上述400例病例上进行测试，并与目前的临床方法进行比较。 该项目的两个关键成果将是：一个史无前例的、精心策划的全身图像数据库， 分段对象和模型系列；自动全身疾病的经过验证的方法 DLBCL的量化和疾病分期。

Quantitative Radiology holds great promise to transform our ability to diagnose, monitor, stage, prognosticate, and detect diseases as well as to plan and guide patient therapeutic interventions. However, the process of locating and delineating anatomic organs and pathologic regions in medical images, known as image segmentation, at a high level of automation has remained a major hurdle to these advances. Most developments on image segmentation have focused on a specific organ or a small group of objects in a specific body region. A new method or a major adaptation of an existing method is engineered when any of these parameters changed. Such an approach is not sustainable and becomes a stumbling block when dealing with whole-body systemic diseases where body-wide image analytics is required. A critical advance is needed in this field to overcome two main challenges: (1) Although prior information about normal anatomy is deemed vital for image segmentation and analysis, its creation and utilization body-wide on a massive scale have not been attempted and are sorely lacking. (2) Techniques to employ such information and methods for body-wide disease quantification at high levels of automation do not exist. The overarching goal is to overcome these challenges by developing a body-wide and generalizable anatomy-guided deep learning image segmentation methodology and demonstrate its application in the study of patients with diffuse large B cell lymphoma (DLBCL) for which PET-based staging and response assessment are of paramount importance. The project has three specific aims. Aim1: To develop a family of body-wide anatomy models representing the entire human adult age spectrum. Existing FDG PET/CT scans of 600 patients from two institutions (Penn and New York Proton Center) covering 10 age groups will be utilized to build anatomy models involving 50 organs and 50 lymph node zones in the extended body torso including neck, thorax, abdomen, and pelvis. A family of 40 anatomy models representing the 4 body regions and 10 age groups will be created from roughly 60,000 3D object samples. Aim2: To develop, implement, and validate a methodology for localizing objects and to quantify disease without explicitly delineating organs and lesions. Gender- and age-specific anatomy models will be utilized for automatically locating the above 100 objects in any given patient PET/CT image and to quantify disease in each body region, organ, and lymph node zone. The methods will be tested on 400 PET/CT images of DLBCL patients. Aim3: To develop and validate an automated method of DLBCL disease staging and prognosis. The disease quantity information will be utilized to develop automated staging and outcome prediction algorithms which will be tested on the above 400 cases in comparison to current clinical methods. Two key outcomes of this project will be: an unprecedented well-curated database of body-wide images, segmented objects, and family of models; and a validated methodology for automatic body-wide disease quantification and disease staging in DLBCL.