Automated Decision Support System for Traumatic Brain Injury through Image Processing and Machine Learning Approaches

摘要： 迫切需要一种用于创伤诊断和预后的自动化决策支持系统 脑损伤(TBI)。脑外伤是现代世界导致死亡的主要原因之一，在很大程度上 伤残和伤残。脑损伤的早期发现及其适当的处理是一项尚未满足的需求。 因此，我们的目标是通过开发一个用于监测的决策支持系统来补充临床医生的决策 以及整合脑外伤患者的可用信息，以进行准确和定量的诊断和预后。 这个项目是一个长期目标的主要组成部分，该目标是建立一个能够创建个性化的系统 治疗计划。具体地说，我们打算自动检测并准确量化两个关键 异常包括大脑中部结构的移位(目标1)和颅内出血(目标2) 计算机断层扫描(CT)头部扫描。在目标1中，我们建立了一个描述大脑空间移位的模型。 结构及其预测能力。我们使用解剖地标来检测3D变形的表面 脑损伤后的脑中线。这种方法使我们能够量化移动量，而这种测量不是 目前是可以实现的。此外，它还提供了对2D中传统中线移动的准确和及时的访问 CT切片。在目标2中，我们建立了一个描述颅内出血及其预测力的模型。我们 实现3D卷积神经网络模型来检测出血区域并进行量化和定位 他们的音量。目前，这些测量是不准确的，并且由于繁琐而不容易获得 手动处理；取而代之的是使用2D CT切片中病变的厚度来评估其严重性。在目标1和 2，我们自动计算传统的和建议的体积和位置测量，并 对它们进行比较，为每个异常提出最佳诊断指标。最后，在目标3中，我们构建了一个 用于脑外伤严重程度评估和结果预测的自动化管道。为此，手动CT扫描显示 将与电子健康记录中提供的患者级别信息相集成，以实现准确 数据驱动的诊断和预后。我们实现了机器学习方法来构建能够 预测短期和长期临床结果。我们的预测模型将独立于我们的 图像处理算法。在实现目标1和目标2后，自动计算来自 CT扫描将被纳入机器学习模型。这项拟议的研究意义重大，因为它 预计将推进脑外伤的治疗，特别是在受伤后的“黄金小时”内。最终，这样的系统 有可能减少延迟和漏诊，从而降低脑损伤的发病率和死亡率。 此外，通过预防永久性和/或继发性伤害，并最大限度地减少住院和 通过康复，我们的系统将有助于减轻美国每年760亿美元的脑损伤护理负担。 除了在提议的方法及其量化输出方面的创新之外，我们汇总了四个现有的 数据集，将表型和治疗的异质性结合在一起，因此得到的模型将是 具有通用性，适用于实际的临床环境。

Summary: There is an urgent need for an automated decision support system for diagnosis and prognosis of traumatic brain injuries (TBI). TBI is one of the leading causes of death in the modern world, and substantially contributes to disability and impairment. The early detection of TBI and its proper management presents an unfilled need. We therefore aim to supplement clinicians' decisions by developing a decision support system for monitoring and integrating available information of a TBI patient for accurate and quantitative diagnosis and prognosis. This project is the main component of a long-term goal of building a system that creates personalized treatment plans. Specifically, we intend to automatically detect and accurately quantify two critical abnormalities including shift in the brain's middle structure (Aim 1) and intracranial hemorrhage (Aim 2) from computed tomography (CT) head scans. In Aim 1, we develop a model for delineating the spatial shift in brain structure and its predictive power. We employ anatomical landmarks to detect a 3D deformed surface of the brain midline after TBI. Such an approach allows us to quantify the shifted volume, a measurement that is not currently achievable. Additionally, it provides accurate and timely access to conventional midline shift in a 2D CT slice. In Aim 2, we build a model for delineating intracranial hemorrhage and its predictive power. We implement a 3D convolutional neural network model to detect hemorrhagic regions and quantify and localize their volume. Currently, these measurements are inaccurate and not readily available due to the cumbersome manual process; instead a lesion's thickness in a 2D CT slice is used to assess its severity. In both Aim 1 and 2, we automatically calculate conventional and proposed volumetric and locational measurements and compare them to suggest the best diagnostic metric for each abnormality. Finally, in Aim 3, we build an automated pipeline for TBI severity assessment and outcome prediction. To this end, manual CT scan reads will be integrated with patient-level information available from electronic health records to achieve accurate data-driven diagnosis and prognosis. We implement machine learning approaches to build models capable of predicting short and long-term clinical outcomes. Our prediction models will be developed independently of our image processing algorithms. Upon achievement of Aims 1 and 2, automatically calculated information from CT scans will be incorporated into machine learning models. The proposed research is significant, because it is expected to advance TBI care, specifically within the “golden hour" post-injury. Ultimately, such a system has the potential to reduce delayed and missed diagnosis, thereby reducing TBI morbidity and mortality. Additionally, by preventing permanent and/or secondary injuries, and minimizing the time of hospitalization and rehabilitation, our system will contribute to reducing the annual $76 billion burden of TBI care in the U.S. In addition to innovation in the proposed approaches and their quantitative outputs, we aggregate four existing datasets to incorporate heterogeneity in both phenotypes and therapies, so the resulted model will be generalizable and applicable to real clinical settings.