可用于第四部分内容

[hyacinth0906]

出处：https://towardsdatascience.com/why-ai-will-not-replace-radiologists-c7736f2c7d80

• 放射医生看片流程

诊断放射学工作流程可以简化为上述可视化的组成步骤: 从病人的表现和病史，导致决定是否进行成像，以及进行何种类型的成像，到安排成像，以及自动化或标准化图像采集。 一旦成像完成，算法将越来越多的后处理图像准备由其他算法解释，在纵向时间框架中登记数据集，提高图像质量，分割解剖和执行检测和量化的生物标志物。 目前，诊断推理似乎是最难解决的难题，而且是人类将保持最多存在的地方。 这将得到智能报告软件、标准化模板和机器可读输出的帮助，使数据能够进行进一步的算法培训，以便更好地为未来的决策软件提供信息。 最后，报告的交流可以通过语言翻译或层面翻译实现半自动化，还可以将结果以有意义的形式呈现给其他临床医生或患者。 这只是开始。 The diagnostic radiology workflow can be simplified into its component steps as visualised above: from patient presentation and history which leads to decision-making on whether or not to image, and what type of imaging to perform, to scheduling the imaging, and automating or standardising image acquisition. Once imaging is done, algorithms will increasingly post-process images ready for interpretation by other algorithms, registering data sets across longitudinal timeframes, improving image quality, segmenting anatomy and performing detection and quantification of biomarkers. At present, diagnostic reasoning seems the toughest nut to crack and is where humans will maintain most presence. This will be aided by the introduction of smart reporting software, standardised templates and machine-readable outputs making data amenable to further algorithmic training to better inform future decision-making software. Finally, communication of the report can be semi-automated via language translation or lay-translation, and augmented presentation of results in a meaningful form to other clinicians or patients can also be accomplished. And this is only for starters…

[hyacinth0906]

https://www.smart-radiology.com/en/

[hyacinth0906]

结构化报告

https://towardsdatascience.com/synoptic-reporting-makes-better-radiologists-and-algorithms-9755f3da511a 对我来说，更重要的是最近人工智能在图像感知方面的推进，它依赖于人类口述的报告，为算法提供数据以便从中学习。 如果人类的自由文本报告已经包含了错误，而且自然语言处理试图解析文本并找到意义，那么这些错误只会被转移。 更糟糕的是，来自人类的免费文本报告完全不同ーー每一位放射科医生的报告都采用不同的风格，使用不同的描述符。 我之前在我的博客里谈到了 CheXnet 的研究，并解释了为什么你不能真正使用人类衍生的自由文本作为一个事实来学习胸部 x 光。 至少不像人类那么准确。 What is more concerning to me is the recent push into artificial intelligence in image perception which relies on human dictated reports to provide data for algorithms to learn from. If human free text reports already contain mistakes, and natural language processing attempts to parse out the text and find meaning, then those mistakes will only be transferred. Even worse, the free text reports from humans are all very different — every radiologist reports in a different style, using different descriptors. I covered this topic before in my blog about the CheXnet study, and explained how you can’t really use human-derived free text as a ground truth to learn what is in chest x-rays. Not as accurately as humans, at least.

我不是唯一关心这件事的人。 来自宾夕法尼亚州好时医疗中心的 Huesch 等人最近发表了他们的研究成果，试图挖掘 CT 扫描寻找肺栓塞(肺动脉血栓)的文字报告。 他们的结论是，自由文本报告"与报告长度和所用报告术语的广泛差异有关"。 作者接着指出,"这些结果支持对结构完整的报告模板以及至少一些强制性的分立领域对报告的易用性及其理解的影响的预期评估"。 换句话说，他们建议报告的结构不仅应该强制标准化，而且应该允许机器学习更好地理解它们。 I’m not the only one concerned about this. Huesch et al, from the Hershey Medical Centre in Pennsylvania recently published their findings on attempting to mine text reports of CT scans looking for pulmonary emboli (blood clots in the arteries of the lungs). They concluded that free text reports were “associated with extensive variability in report length and report terms used”. The authors go on to state that “these results support the prospective assessment of the impact of a fully structured report template with at least some mandatory discrete fields on ease of use of reports and their understanding”. In other words, they are suggesting that reports should be structured not only to force standardisation, but also to better allow machine learning to understand them.

显然，我们必须减少误差，使医学数据可以被人和机器理解。 随着放射学标准化词典的兴起，以及将结构化报告整合成 DICOM 格式的工具，我们看到技术与临床需求的融合，最终可能使结构化报告成为一种可用的现实，通过所谓的'概要报告'来实现。 这是直接从输入数据生成机器可读报告的概念。 通过这种方式，临床数据在源头被标记，附加到任何你喜欢的编码系统(ICD-10，FIHR，SNOMED) ，然后在一个结构化的模板中处理成一个自由的文本报告。 底层的编码数据比传统的自由文本报告更适合于计算分析，为准确大规模查询放射学数据打开了大门。 It is clear that we have to reduce error and make medical data understandable by both man and machine. With the rise of a standardised lexicon for radiology, and tools for integrating structured reports into DICOM format, we are seeing a convergence of technology with clinical need that could finally make structured reporting a usable reality, enabled by what is known as ‘synoptic reporting’. This is the concept of producing machine-readable reports directly from input data. In this manner, clinical data is tokenised at source, attached to whatever coding system you like (ICD-10, FIHR, SNOMED), and then processed into a free text report within a structured template. The underlying coded data is then much better suited to computational analysis than the traditional free text-only report, opening up the doors to accurate large scale interrogation of radiology data.

在2017年的 RSNA 上，一家供应商 Smart Reporting 为用户呈现了一个优雅的概要报表解决方案。 该软件将数据字段转换成人类可读的自由文本，而不是说教式的打勾选择。 使用语音命令(与领先的放射科语音听写软件 Nuance 链接) ，放射科医生只需通过说话来导航模板字段，就不需要点击鼠标了。 At RSNA 2017, one vendor, Smart Reporting showcased an elegant solution for synoptic reporting. Instead of didactic ticking of boxes, the software converts data fields into human-readable free text. Using voice command (linked in with Nuance, the leading radiology voice-dictation software), radiologists can just speak to navigate the template fields, eliminating the need for mouse clicks.

每个领域都是完全可定制的，因此每个放射科都可以根据需要设置自己的格式，或者为每个临床用例选择使用次专业专家设计的模板。 随着放射科医生的加入，实时决策支持被显示出来(如上面关于肺结节测量的例子) ，确保放射科医生的表现都达到同样的金标准。 正常的研究可以在几秒钟内以你的部门和参考临床医生的方式报告。 癌症研究尤其受益于详细的标准化，而智能报告的目的是使填写这些类型的报告更少的杂务。 Each field is entirely customisable, so every radiology department can set out their own format as required, or opt to use templates designed by sub-speciality experts for each clinical use-case. As the radiologist goes along, real-time decision support is shown (as in the example above on lung nodule measurements), ensuring that radiologists all perform to the same gold standard. ‘Normal’ studies can be reported in a matter of seconds in a format your department and referring clinicians like. Cancer studies in particular benefit from detailed standardisation, and Smart Reporting aims to make filling out those types of reports less of a chore.

对于那些担心正常的机智的报告失去细微差别的人，不要担心，你可以编辑一个模板，包括你最喜欢的评论，因为你知道你的报告仍然有充分的结构。 你编辑的基础数据点仍然是一样的，但是文本是你自己的。 更重要的是，你的俏皮话可以立即翻译成多种语言，甚至还有可能从同样的结构化数据中创造出面向病人的报告，这是我在放射学中用于人工智能的五大用例之一。 For those of you concerned about losing the nuance of your normal witty reports, don’t worry, you can always edit a template to include your favourite sayings, safe in the knowledge that your report is still adequately structured. The underlying data point that you edit remains the same, but the text is your own. Even more, your witticisms can be instantly translated into multiple languages, and there is even the potential to create patient-facing reports from the same underlying structured data, one of my top five use-cases for artificial intelligence in radiology.

另一个聪明的特点是，你可以将人工智能图像感知算法的输出，自动地将相应的文本(编辑成您的喜好)放入报表中。 目前正在荷兰 Radboud 大学医学中心与索诺纳成像分析中心进行试验。 还有一个集成的分析套件，这样你就可以深入了解所有报告的数据，这是仅仅使用免费文本是不可能的。 所有这些数据点都可以潜在地与医院的电子记录联系起来，用于审计、数据分析和机器学习研究。 这与当前的热门话题以及卫生技术中更为热门的"真实世界数据"业务。 那些知道如何在规模上创造和利用准确的现实世界数据的人正在看到大量的回报; 只要看看弗拉蒂龙，今年罗氏以19亿美元的价格收购了它。 创建和管理自己的数据湖的价值是显而易见的。 现在，放射科医生应该接受这一点。

[hyacinth0906]

可用于人工智能的标准数据并不容易

The problem is … medical imaging data isn’t ready for AI 问题是... 医学成像数据还没有为人工智能做好准备

级别 d 的数据在 Level D data is that which is unverified in quantity and quality, is inaccessible and is in a format which makes it difficult or impossible for machine learning to do anything. This level of un-anonymised data is in every single hospital trust PACS archive in massive volumes, just sitting there, doing nothing, except acting as a record of clinical activity. (And…I shudder at the thought… every so often due to data storage issues, NHS trusts actually delete backlogs of this data. Like throwing away oil…)数量和质量上未经验证，是不可访问的，其格式使机器学习做任何事情都变得困难或不可能。 这种水平的匿名数据在每个单独的医院信托 PACS 档案库中都有大量的，只是坐在那里，什么也不做，除了充当临床活动的记录。 (而且... ... 我对这种想法感到不寒而栗... ... 偶尔由于数据存储问题，NHS 实际上信任删除这些数据的积压。 就像扔掉石油... ...)

为了将 d 级数据提炼到 c 级，您需要建立一个精炼厂。 数据精炼的第一阶段是获得对你的数据访问的道德许可。 实际上，这是通过一个数据共享协议来实现的，可以通过一个道德委员会在本地与自己达成，也可以通过第三方(大学、公司或初创企业)达成。 国家医疗服务系统的信托可能有数以千计的数据共享协议在任何地方在同一时间。 这些协议还将包括有关匿名数据的条款，因为显然没有人希望 NHS 泄露病人的机密信息。 到目前为止还不错... 但是，数据仍然是非结构化的，并不能代表一个完整的集合。 它也将是非常嘈杂的，充满了错误，遗漏和只是普通的奇怪的条目。 现在，不管是谁访问了这些数据，都必须先弄清楚如何让它变得有用，然后才能破解算法开发。 C 级别的数据已经可以提供给人工智能开发者了，但是还远远没有达到有用的程度。 In order for Level D data to be distilled up into Level C you need to build a refinery. The first stage of data refining is to get ethical clearance for access to your data. In practice, this is done through a data sharing agreement either locally with yourself via an ethics committee, or with a third party (university, company or start-up). NHS trusts may have thousands of data sharing agreements in place at any one time. These agreements will also include clauses on data anonymisation, as obviously no-one wants the NHS to give away confidential patient information. So far so good…however, the data is still very unstructured, and will not be representative of a complete set. It will also be extremely noisy, full of errors, omissions and just plain weird entries. Whoever has access to the data now has to figure out how to make it useful before they can get cracking on algorithmic development. C level data is ready to be given out to AI developers, but is still far from ready for being useful.

The data now needs to be refined further into Level B data, by structuring it, making sure it’s representative of the data you think you have, and running visualisations on it to be able to draw out insight into noise characteristics and other analysis metrics. This process is actually even harder than the D to C stage, as it is bespoke for each dataset. There is no standard way of checking medical imaging data, and each individual group that has access to your data will be running their own data visualisation and analyses. This is because data from different hospitals will have different characteristics and be in different formats (e.g different DICOM headers, date and time stamps etc). The process of converting Level C to B can take months — not exactly what researchers or start-ups need in the race for gold. Only with Level B data can you have an idea of what is possible with it, and where AI can be used to solve actual problems. 现在需要将数据进一步细化为 b 级数据，通过构造它，确保它代表你认为你拥有的数据，并在它上面运行可视化，以便能够洞察噪音特性和其他分析指标。 这个过程实际上比 d 到 c 阶段更难，因为每个数据集都是定制的。 没有标准的方法来检查医疗成像数据，每个单独的群体都可以接触到你的数据，他们将运行他们自己的数据可视化和分析。 这是因为来自不同医院的数据具有不同的特征和不同的格式(例如不同的 DICOM 头部、日期和时间戳等)。 将 c 级转化为 b 级的过程可能需要几个月的时间，这并不是研究人员或初创企业在淘金竞赛中所需要的。 只有使用 b 级数据，你才能知道什么是可能的，什么地方可以用人工智能来解决实际问题。

Level A data is that which is as close to perfect for algorithmic development — it is structured, fully annotated, has minimal noise and, most importantly, is contextually appropriate and ready for a specific machine learning task. An example would be a complete data set of 1million liver ultrasounds with patient age, gender, fibrosis score, biopsy results, Liver Function Tests (LFTs), and diagnosis all structured under the same meta-tags, ready for a deep learning algorithm to figure out which patients are at risk of Non-Alcoholic Fatty Liver Disease (NAFLD) on B-mode US scans. A 级数据是一种近乎完美的算法开发的数据，它是结构化的，完全注释的，具有最小的噪音，最重要的是，它是适合于具体的机器学习任务的。 一个例子是一套包含1百万个肝超声波的数据集，其中包括患者年龄、性别、纤维化评分、活检结果、肝功能检查(LFTs) ，以及所有结构在相同的 meta 标签下的诊断，为深度学习算法做准备，以便找出哪些患者在 b 型超声波扫描中面临非酒精性脂肪肝疾病(NAFLD)的风险。

Annotation is perhaps the hardest part in radiology dataset refining — each and every image finding should ideally be annotated by an experienced radiologist so that all possible pathologies are highlighted accurately and consistently across the entire data set. The problem is — barely any of the existing medical imaging data anywhere in the world is annotated in this way. In fact, most images in the ‘wild’ aren’t even annotated. That’s why there is an entire industry centred around data tagging. Ever signed into a website and been asked to click on images that contain road signs or cars? You’re tagging data for self-driving car algorithms! Of course, not every internet user is a radiologist, so this crowdsourcing model doesn’t work for medical imaging (unless you run Radiopaedia — there’s a free business idea for you guys!). Instead, researchers have to beg or pay for radiologist time to annotate their datasets — a monumentally slow and expensive task (trust me, I spent 6 months during my thesis drawing outlines around prostates…). The alternative is to use Natural Language Processing (NLP) on imaging reports to word-mine concepts and use them to tag images — however, this model is far from proven to be robust enough (yet). 注释也许是放射数据集中最难的部分ーー理想情况下，每一张图像发现都应该由一位经验丰富的放射科医生加以注释，以便所有可能的病理都能在整个数据集中准确而一致地突出显示。 问题在于，世界上几乎没有任何现有的医学影像数据是以这种方式注释的。 事实上，大多数野生的图片甚至没有注释。 这就是为什么整个行业都以数据标签为中心。 曾经登陆过一个网站，并被要求点击包含路标或汽车的图片吗？ 你正在给数据打上标签，用于自动驾驶汽车 / 值算法！ 当然，并不是每个互联网用户都是放射科医生，所以这种群众外包模式并不适用于医疗成像(除非你运营 Radiopaedia ー你们有免费的商业想法!) . 相反，研究人员不得不乞求或支付放射学家的时间来注释他们的数据集——这是一项极其缓慢和昂贵的任务(相信我，我花了6个月的时间在我的论文中绘制关于前列腺的大纲...)。 另一种方法是使用自然语言处理(NLP)对文字挖掘概念的图像报告，并使用它们标记图像ー然而，这个模型还远远没有被证明是足够的(尚)。