BIGDATA Small Project Structurization and Direct Search of Medical Image Data

描述(申请人提供)：IBM估计，世界上30%的全部数据是医疗信息。医学图像占据了医疗记录的很大一部分，在美国大约有1亿次扫描，而且每年都在增长。此外，每次扫描的数据大小随着图像分辨率的提高而稳步增加。这些大数据不是结构化的，由于缺乏标准化的成像协议，它们具有高度的异构性，具有不同的空间分辨率、对比度、切片方向等。在本项目中，我们将开发一种结构化和搜索医学成像信息的技术，这将使过去的数据可用于教育和循证临床决策。在这笔赠款中，我们将重点放在脑部核磁共振上，这是核磁共振数据中最大的部分。目标群体将是做出决定的医生，患者将是最终受益者。目前，放射影像数据存储在称为PACS的临床数据库中。PACS中的图像数据不是结构化的。因此，一旦患者的诊断完成，PACS中的大多数数据目前都会在档案中被丢弃。放射科医生依靠他们的经验和教育来做出医疗决定。这是医学实践中的一个典型问题，需要客观的循证医学。有许多正在进行的尝试来构建PACS的文本字段，其中包括自由文本放射学报告、临床信息和诊断的自然语言处理。在我们的方法中，我们建议结构化图像数据，而不是文本字段，以支持对图像的直接搜索。也就是说，医生将提交一张新患者的图像，并搜索具有相似解剖表型的过去图像。然后，将检索到的数据的临床报告汇编成诊断和预后的统计报告。我们相信，这种图像结构化是“解锁PACS中目前存储的海量信息”并将其用于教育和现代循证医学决策的关键。具体目标是：目标1：开发和测试高通量图像结构化技术的准确性目标2：开发和测试图像搜索引擎目标3：能力建设要求：为研究和教育目的开发用于数据结构化/搜索服务的原型云系统

DESCRIPTION (provided by applicant): IBM estimates that 30% of the entire data in the world is medical information. Medical images occupy a significant portion of medical records with approximately 100 million scans in US and growing every year. In addition, the data size from each scan steadfastly increases as the image resolution improves. These BigData are not structured and due to lack of standardized imaging protocols, they are highly heterogeneous with different spatial resolutions, contrasts, slice orientations, etc. In this project, we will deelop a technology to structure and search medical imaging information, which will make the past data available for education and evidence-based clinical decision-making. In this grant, we will focus on brain MRI, which comprises the largest portion of MRI data. The target community will be physicians who make decisions and the patients will be the ultimate beneficiaries. Currently, radiological image data are stored in clinical database called PACS. The image data in PACS are not structured. Consequently, once the diagnosis of a patient is completed, most of the data in PACS are currently discarded in the archive. Radiologists rely on their experience and education to reach medical decisions. This is a typical problem in medical practice that calls for objective evidence-based medicine. There are many ongoing attempts to structure the text fields of PACS, which include natural language processing of free-text radiological reports, clinical information, and diagnosis. In our approach, we propose to structure the image data, not text fields, to support direct search of images. Namely, physicians will submit an image of a new patient and search past images with similar anatomical phenotypes. Then, the clinical reports of the retrieved data will be compiled for a statistical report of the diagnosis and prognosis. We believe this image structuration is the key to "unlock the vast amounts of information currently stored" in PACS and use them for education and modern evidence-based medical decisions. The specific aims are; Objective 1: To develop and test the accuracy of high-throughput image structuration technologies Objective 2: To develop and test the image search engine Objective 3: Capacity Building Requirement: To develop prototype cloud system for data structuration / search services for research and educational purposes