Artificial Intelligence and the implications on Medical Imaging

• 分析与建模 - 机器学习

• 医疗保健和医院

• 自动化疾病诊断

有几个因素同时推动人工智能在放射学中的整合。首先，在世界上许多国家，接受过放射学培训的医生数量与对诊断成像的需求不断增长之间存在差异。这导致对工作效率和生产力的更高要求。例如，2012 年至 2015 年期间，英格兰放射科专家（顾问劳动力）的数量增加了 5%，而同期 CT 和 MR 扫描的数量分别增加了 29 个和 26 个百分点。在苏格兰，差距进一步扩大（皇家放射科医师学院，2016 年）。今天，放射科医生平均每 3 到 4 秒解读一次图像，每天 8 小时（Choi 等人，2016 年）。

其次，当今扫描仪的图像分辨率不断提高——导致数据量越来越大。事实上，估计的整体医疗数据量每三年翻一番，这使得放射科医生越来越难以在没有计算机数字处理的额外帮助的情况下充分利用可用信息。在放射学研究和临床诊断中，希望能够定量分析这些大量未开发的数据，例如，利用新的可测量成像生物标志物来评估疾病进展和预后（O'Connor et al. 2017） .专家们看到放射学从定性解释学科转变为定量分析学科的巨大未来潜力，定量分析从广泛的数据集（“放射组学”）中获取临床相关信息。 “图像不仅仅是图片，它们是数据，”美国放射科医生 Robert Gillies 和他的同事写道（Gillies et al. 2016）。当然，放射学的这个方向将需要强大的自动化程序，其中一些至少将属于人工智能领域。

在医学成像中使用机器学习并不新鲜——然而，今天的算法比传统应用程序强大得多（van Ginneken 2017）。深度学习所基于的人工神经网络总是有多个功能层，有时甚至超过一百个，可以包含数千个神经元和数百万个连接。 （例如，只有一个中间层的简单人工神经网络被描述为“浅层”网络。）所有这些连接都在人工神经网络的训练过程中通过它们各自参数的逐渐变化来调整——用数学术语来说：它们的权重。通过这种方式，深度网络具有几乎无法想象的处理信息的可能组合数量，甚至可以对高度复杂的非线性上下文进行建模。在训练过程中，人工神经网络的不同层越来越多地使用每个连续层来构建输入数据，从而对信息产生更抽象的“理解”。当然，只有通过先进的数学方法以及更高的计算能力和更快的图形处理器（GPU）来计算学习过程中的无数步骤，才能使这种深度人工神经网络成为可能。 2013 年，《麻省理工科技评论》将深度学习确定为年度十大突破性技术之一（Hof 2013）。

对于图像识别，“深度卷积神经网络”（一种特定类型的 ANN）已被证明特别有效。与大脑中的视觉皮层类似，这些网络首先从输入数据中提取基本图像特征，如角落、边缘和阴影。然后在多个抽象步骤中，他们独立解决更复杂的图像模式和对象。当这些网络中最好的网络在非医学图像数据库上进行测试时，它们的错误率现在下降到只有百分之几（He et al. 2015）。此外，可以结合不同的网络架构和方法（例如，深度学习与“强化学习”）以根据所提出的问题获得最佳结果。

鉴于这一发展，专家预计医学成像将发生重大变化（Lee 等人，2017 年）。与 1990 年代后期开始在美国引入的以前的 AI 方法不同，特别是用于乳房 X 线照相筛查，但存在很多缺点（Morton 等人 2006；Fenton 等人 2007；Lehman 等人 2015），今天的算法可能会被证明是临床诊断的变革性技术。

• The promise of AI in medical imaging lies not only in higher automation, productivity and standardization, but also in an unprecedented use of quantitative data beyond the limits of human cognition. This will support better, and more personalized, diagnostics and therapies. Today, artificial intelligence already plays an important role in the everyday practice of image acquisition, processing and interpretation. Siemens Healthineers, for example, has developed a pattern recognition algorithm (Automatic Landmarking and Parsing of Human Anatomy, ALPHA) for its 3D diagnostic software “syngo.via,” which automatically detects anatomical structures, independently numbers vertebrae and ribs, and also aids in precisely overlaying different exami- nation dates or even different modalities (AI-based landmark detection and image registration). This considerably helps simplify workflows in diagnostic imaging. The same is true for award- winning algorithms like “CT Bone Reading” for virtual unfolding (2D reformatting) of the rib cage or “eSie Valve” for simultaneous 3D visualization of heart valve anatomy and blood flow (R&D Magazine 2014; R&D 100 Conference 2015). AI applications like these are already an established part of available imaging software.

• Numerous other applications are in development (Comaniciu et al. 2016) and can be expected from a range of companies in coming years (Signify Research 2017). The corporate research of Siemens Healthineers alone includes 400 patents and patent applications in the field of machine learning, 75 of them in deep learning.

  No less significant is the availability of comprehensive open-source tools for developing AI applications (Erickson et al. 2017). And many academic research groups are moving clinical implementa- tions of machine methods forward through pilot studies. Realistic scenarios for routine clinical use of AI might include, for example, improved assessments of chest ultrasound images and detection of pulmonary nodes in CT (Cheng et al. 2016), or quantitative analyses of neurological diseases through precise segmentation of brain structures (Akkus et al. 2017).

• Intelligent algorithms would also benefit cardiac patients undergoing a coronary CT angiogram, since deep learning methods can be used to calculate the calcium score of their vessels at the same time. Until now, an additional CT scan is often performed for this purpose, with added radiation exposure (Wolterink et al. 2016).

  Last but not least, the use of artificial intelligence offers remarkable prospects for countries with fewer medical resources. A recent study has shown that tuberculosis of the lung can be detected on chest x-rays with 97% sensitivity and 100% specificity, if the images are analyzed by two different deep ANNs, and only those cases in which the algorithms do not concur are then evaluated by a doctor trained in radiology. Such a workflow could have great practical relevance in regions with widespread tuberculosis, but few radiologists on hand (Lakhani & Sundaram 2017).

• The acceleration of certain work steps in diagnostic imaging with artificial intelligence already is a reality today. For example, AI algorithms enable automated detection of anatomical structures, intelligent image registration and reformatting. These kind of efficiency gains will become increas- ingly important given the growing demand for diagnostic imaging and rising cost pressure.

  In the longer term, AI-based image analyses with reproducible characteristic measurements (imaging biomarkers), indices and “lab-like” results will likely prevail, particularly in areas like cardiac imaging that are already quantitatively oriented. This will also favor the (semi-)automated drafting of radiology reports and the transformation of radiology to a data-driven research discipline (“radiomics”). Meanwhile, radiological data sets can not only be analyzed graphically
  by artificial intelligence, but they can also be annotated textually (Shin et al. 2016).

• The implementation of AI offers particularly fascinating prospects for personalized diagnostics and treatment. Through interoperable information systems, for example, clinical patient data could be linked with imaging algorithms to pursue individualized scanning strategies. In addition, AI applications would likely enable more precise diagnoses and more meaningful risk scores by gathering together large quantities of information. As a large multicenter study recently showed, for example, the long-term risk of mortality of patients with suspected cardiovascular diseases can be estimated with much greater precision if manifold clinical and CT angiogram parameters are integrated into a personalized prognosis model using machine learning procedures (Motwani et al. 2017). Such AI-based approaches could in future better identify high-risk patients, but also help prevent unnecessary treatments, and thus involve diagnostic radiology more closely in outcome- oriented clinical decisions.