There’s a shortage in our healthcare system.

我们的医疗保健系统短缺。

A recent finding from the Association of American Medical Colleges found that by 2032 we will be short between 46,900 and 121,900 physicians just in the United States and it isn’t much better in other countries.

美国医学院联合会的一项最新发现发现，到2032年，仅在美国，我们的医生人数将不足46,900至121,900 ，而在其他国家，情况并没有好得多。

This means physicians will need to work more hours, with more work on their plates, while trying to maintain quality care for their patients.

这意味着医师将需要花费更多的时间进行更多的工作 ，同时还要努力保持对病人的高质量护理 。

We’ve seen just how fragile our healthcare system is with the recent pandemic and even ignoring that physician burnout is on the rise.

我们已经看到随着最近的大流行，我们的医疗体系多么脆弱，甚至忽略了医师倦怠的趋势在加剧 。

“Researchers surveyed 422 family physicians and general internists who worked in 119 ambulatory care clinics and surveyed 1,795 patients from these clinics and reviewed their medical records for information on care quality and medical errors. More than half of the physicians reported experiencing time pressures when conducting physical examinations. Nearly a third felt they needed at least 50 percent more time than was allotted for this patient care function. In addition, nearly a quarter said they needed at least 50 percent more time for follow-up appointments.” — Agency for Healthcare Research and Quality

研究人员调查了在119个门诊诊所工作的422位家庭医生和普通内科医师，并调查了这些诊所的1,795名患者，并查看了他们的病历以获取有关护理质量和医疗错误的信息。 超过一半的医师报告进行身体检查时承受时间压力。 将近三分之一的人认为，他们需要的时间 比分配给该患者护理功能的 时间至少多50％ 。 此外，将近四分之一的人表示，他们至少需要多花50％的时间进行随访。 ” — 医疗保健研究与质量局

But we can build systems to help reduce the burden on physicians, by streamlining their workflows and giving them access to quality tools to make it easier for them to their jobs.

但是，我们可以构建系统来简化医生的工作流程，并为他们提供优质工具的访问权限 ， 从而减轻他们的负担，从而使他们的工作更轻松。

This article will serve as an introduction to the application of AI in the healthcare system that I will be writing. Here I will be focusing on clinical imaging and the application of AI algorithms to aid clinical physicians.

本文将作为我将要编写的AI在医疗保健系统中的应用的介绍。 在这里，我将重点介绍临床成像以及AI算法在协助临床医生中的应用。

用于临床成像的工具 (Tools Used For Clinical Imaging)

When training a machine learning model, it's important to have a general understanding of the kinds of images your dataset may contain. Some images may be very detailed and allow for higher training accuracies. Other images may have lower detail but are used more frequently.

在训练机器学习模型时，对数据集可能包含的图像类型有一个总体了解很重要。 一些图像可能非常详细，可以提高培训的准确性。 其他图像的细节可能较低，但使用频率更高。

For example, you may develop a great model that can determine where bone fractures are on a CT scan but the model may perform poorly on x-ray scans. If most hospitals and clinics use x-rays for bone fractures and your model isn’t able to perform well on that kind of data then it might be useless.

例如，您可能开发了一个很好的模型，该模型可以确定CT扫描中的骨折部位，但是该模型在X射线扫描中的表现可能不佳。 如果大多数医院和诊所使用X射线检查骨折，而您的模型在这种数据上表现不佳，那么它可能就没有用了。

That’s why it’s important to know the types of imagining tools available so you can research how to best implement the model you build.

因此，重要的是要知道可用的成像工具的类型，以便您可以研究如何最好地实现所构建的模型。

There are three kinds of machines that we’ll focus on that are used by clinicians.

临床医生将使用三种机器。

1. X射线 (1. X Rays)

X-rays are typically used for bones and sometimes dense organs like the lungs or the heart.

X射线通常用于骨骼，有时用于密集的器官，如肺或心脏。

The machine works by passing radiation (x-ray) through an area of the body. Radiation that gets blocked by dense objects appears white in the film.

机器通过使辐射(X射线)穿过身体的某个区域来工作。 被稠密物体阻挡的辐射在胶片中显示为白色。

They’re typically used to diagnose :

它们通常用于诊断：

Disease / Bone Degradation

疾病/骨骼退化 Discoloration

变色 Fractures

骨折 Tumors

肿瘤 Infections

传染病

These images are usually less detailed than other imaging solutions and require radiation to be used.

这些图像通常不如其他成像解决方案详细，并且需要使用辐射。

Wikipedia 维基百科

The x-ray will capture a single 2D image.

X射线将捕获单个2D图像。

X-Rays are usually the cheapest option for medical imaging and are sometimes done first even if a clinician thinks you may need a CT or MRI afterward.

X射线通常是用于医学成像的最便宜的选择，即使临床医生认为之后可能需要CT或MRI，有时也要先进行 X射线检查。

2. CT扫描 (2. CT Scans)

A CT scan is similar to an x-ray in that is uses radiation to capture an image. The main difference is that it has the ability to capture multiple slices of the body, giving the physician a 3D view of the section being captured.

CT扫描类似于X射线，它使用辐射来捕获图像。 主要区别在于它具有捕获身体多个切片的能力，从而为医生提供了所捕获部分的3D视图。

Wikipedia 维基百科

CTs are very powerful because they allow clinicians to create detailed images of bone, blood vessels soft tissues, and other organs.

CT非常强大，因为它们允许临床医生创建骨骼，血管，软组织和其他器官的详细图像。

They’re typically used to diagnose :

它们通常用于诊断：

Appendicitis

阑尾炎 Cancer

癌症 Trauma

外伤 Heart Disease

心脏病 Infections.

感染。

It’s important to note the image itself is not 3D, but multiple 2D images.

重要的是要注意图像本身不是3D 而是多个2D图像 。

CT scans are more expensive, than x-rays but provide much more detail.

与X射线相比，CT扫描更昂贵，但可以提供更多细节。

核磁共振扫描 (MRI Scans)

Magnetic Resonance Imaging (or MRI) is usually used for soft tissue injuries like muscles or connective tissues. This is because they have provided a greater amount of detail for these types of tissues. MRI’s use an electromagnet to produce an image so they are free from radiation.

磁共振成像(或MRI)通常用于诸如肌肉或结缔组织的软组织损伤。 这是因为它们为这些类型的组织提供了更多细节。 MRI使用电磁体产生图像，因此它们没有辐射。

However, this means that they cannot be used by people with metal in their bodies. MRI’s can produce both 2D and 3D images

但是，这意味着它们不能被带金属的人使用。 MRI可以生成2D和3D图像

Wikipedia 维基百科

MRI’s are the most expensive imaging tool a physician may have.

MRI是医生可能拥有的最昂贵的成像工具。

Source: You can read more about differences here

资料来源： 您可以 在此处 详细了解差异

医学影像工作流程 (The workflow in medical imaging)

The next part that’s important to understand is the workflow that clinics and hospitals generally take to imaging.

下一个需要理解的重要部分是诊所和医院通常用于成像的工作流程。

This information will help us know, at what stage our model should be deployed.

这些信息将帮助我们知道应在什么阶段部署模型。

Here we see 5 main stages. The image is first taken by using an X-ray, CT, or MRI. The images are then stored in an online server called the Picture Archiving and Communication System (PACS). This server allows different departments to gain access to the imaging information in a hospital. A radiologist will then read and interpret the images in a first-in first-out order. They will identify abnormalities and then generate an interpretation of the image or images of a patient. Finally, a physician will take the results of the scan and all other available information to make a diagnosis. It’s important to know that the diagnosis is only made by the physician and not the radiologist.

在这里，我们看到5个主要阶段。 首先使用X射线，CT或MRI拍摄图像。 然后将图像存储在称为图片存档和通信系统(PACS)的在线服务器中。 该服务器允许不同部门访问医院中的成像信息。 放射线医师将按照先进先出的顺序读取和解释图像。 他们将识别异常，然后生成患者图像的解释 。 最后， 医生将获取扫描结果和所有其他可用信息以进行诊断。 重要的是要知道诊断仅由医师而不是放射线医师做出。

成像类型 (Types of Imaging performed)

1.诊断成像 (1. Diagnostic Imaging)

If a clinician believes that a patient has an issue of some sort and imagining is required to verify it then we would call this a diagnostic image.

如果临床医生认为患者存在某种问题，并且需要想象以对其进行验证，那么我们将其称为诊断图像。

For example, in the case where a physician may believe that there was severe damage to the brain from a concussion, they may order a CT or MRI scan to diagnose if there was any significant brain damage.

例如，在医生可能认为脑震荡对大脑造成严重损害的情况下，他们可能会命令CT或MRI扫描以诊断是否存在明显的脑部损害。

Diagnostic Images can be used in both emergency situations where a life may be at risk (for example verifying a brain bleed) or in non-emergency situations.

诊断图像可用于可能存在生命危险的紧急情况(例如，验证脑出血)或非紧急情况下。

2.筛选 (2. Screening)

This is typically used when there’s nothing wrong with the patient necessarily but they’re non the less high risk. For example, a patient with a family history of lung cancer may receive periodic screening for early detection.

通常在患者没有任何问题但风险不高的情况下使用。 例如，有肺癌家族史的患者可能会接受定期筛查以进行早期检测。

These are typically non-emergency situations.

这些通常是非紧急情况。

Source: Mayfair

资料来源 ： 梅费尔

基本的2D图像算法 (Basic 2D Image Algorithms)

Next, we should get a brief overview of the types of imaging algorithms that can be applied to medical images. These algorithms will help us determine the best approach and solution to solving specific problems.

接下来，我们将简要概述可应用于医学图像的成像算法的类型。 这些算法将帮助我们确定解决特定问题的最佳方法和解决方案。

分类 (Classification)

Classification is used to identify what class an image falls into. This is done by the algorithm either identifying a structure or finding with an image. This can be binary or multiclass. In laymen terms, classification solves the “is this a cat or dog photo”. In medicine, we can use it to detect if something like if a tumor is present in an image.

分类用于识别图像所属的类别。 这是通过算法完成的，即识别结构或查找图像。 这可以是二进制或多类。 用外行术语来说，分类解决了“这是猫还是狗的照片”。 在医学上，我们可以使用它来检测图像中是否存在肿瘤。

For example, if I were to give this to an image classifier, it would return we with the result of “Patient has tumor”. It would not provide any other information, like the size or location of the tumor.

例如，如果我将其提供给图像分类器，它将以“ 患者患有肿瘤 ”的结果返回给我们。 它不会提供任何其他信息，例如肿瘤的大小或位置。

Wikipedia 维基百科

本地化(检测) (Localization (Detection))

Localization is similar to classification in that it will identify whether a structure or finding exists. But it goes a little further.

本地化与分类相似，因为它将识别是否存在结构或发现。 但它走得更远。

Instead of telling you that a finding exists, localization will box-in the locations of each of the findings that the model had identified. This is extremely useful in drawing in the attention of a radiologist who can more quickly identify potential points of interest.

本地化不会告诉您发现的存在，而是会在模型中标识出每个发现的位置。 这在引起放射科医生的注意时非常有用，放射科医生可以更快地识别潜在的兴趣点。

In the above example, the tumor is glaringly obvious. But in the next example, we’ll be looking for pulmonary lesions.

在以上示例中，肿瘤非常明显。 但是在下一个示例中，我们将寻找肺部病变。

Median Technologies 中位数技术

Here, the algorithm was able to correctly identify the pulmonary lesion. Sometimes a confidence level is also given to help the radiologist understand how likely the model believes that it identified the correct area.

在这里，该算法能够正确识别肺部病变。 有时还会给出置信度，以帮助放射线医师了解模型认为其识别出正确区域的可能性。

分割 (Segmentation)

Segmentation is similar to localization and classification in that it will identify whether a structure or finding exists and locate their positions on an image.

分割类似于定位和分类，因为它将识别结构或发现是否存在，并在图像上定位它们的位置。

Segmentation will identify the pixels in an image that contain a structure and highlight that structure. This is useful for obtaining the area or size of a structure and tracking its growth or shrinking over time.

分割将识别图像中包含结构的像素并突出显示该结构。 这对于获取结构的面积或大小以及跟踪其随时间的增长或收缩很有用。

For example, here in this example, instead of just boxing the area where the tumor is, the model instead tries to outline the exact tumor. This makes it easier to computationally determine the size of a tumor from a certain slice or image.

例如，在此示例中，模型不仅试图将肿瘤所在的区域装箱，而且还会尝试勾勒出确切的肿瘤轮廓。 这使得从某个切片或图像计算确定肿瘤大小变得更加容易。

ingegneriabiomedica.org ingegneriabiomedica.org

利益相关者 (Stakeholders)

If you want to be able to bring your product to market, it's important to note which stakeholders are involved and forming your product in such a way as to benefit each of the stakeholders.

如果您希望能够将产品推向市场，那么重要的是要注意涉及哪些利益相关者，并以使每个利益相关者都受益的方式构成您的产品。

放射科医生(临床利益相关者) (Radiologists (Clinical Stakeholder))

This will usually be the end-user of your algorithm and should be familiar with the accuracy and the types of failures the algorithm may make.

这通常是您算法的最终用户，并且应该熟悉算法的准确性和失败类型。 They will play an important advisory role when developing your model, providing important insights that they face on a day-to-day basis.

在开发模型时，他们将扮演重要的咨询角色，提供他们每天面对的重要见解。 Radiologists care most about having low distribution to there already busy workflow. Having a great performing model that is difficult to use won’t fly.

放射科医生最关心的是对已经很忙的工作流程分配得很少。 拥有难以使用的出色性能模型将不会成功。

临床医生(临床利益相关者) (The clinician (Clinical Stakeholder))

Clinicians will have less visibility into the understanding of your algorithm.

对于您对算法的理解，临床医生的可见度会降低。 Clinicians also care about low disruption to there workflow. An algorithm that takes too long or provides data that isn’t easy for a clinical to interpret will be useless to them.

临床医生还关心那里的工作流程中断率低。 花费太长时间或提供临床难以解释的数据的算法对他们将毫无用处。

患者(临床利益相关者) (Patients (Clinical Stakeholder))

The patient will be the most important stakeholder.

病人将是最重要的利益相关者。 The patient may never know that an AI was involved in the process of there diagnosis.

患者可能永远不知道AI参与了诊断过程。 The patient cares most about understanding why their diagnosis is the way it is, the length of time to obtain a diagnosis, and the accuracy of this diagnosis.

病人最关心的是理解为什么要这样诊断，获得诊断的时间长短以及诊断的准确性。

医疗器械公司(行业利益相关者) (Medical Device Companies (Industry Stakeholders))

These companies manufacture the devices used in hospitals, labs, and clinics.

这些公司制造用于医院，实验室和诊所的设备。 They usually have their own accompanying imaging software and develop their own AI algorithms.

他们通常拥有自己的随附成像软件并开发自己的AI算法。 Medical device companies

医疗器械公司 It’s important to know what these systems are capable of and how your algorithm improves or enhances the current software already used

了解这些系统的功能以及您的算法如何改进或增强已使用的当前软件非常重要

软件公司(行业利益相关者) (Software Companies (Industry Stakeholders))

Develop AI algorithms for imaging or clinical applications.

开发用于成像或临床应用的AI算法。 They have to be explicit about which types of medical imagining hardware their software has been validated on (for regulatory purposes).

他们必须明确说明其软件已在哪种类型的医学成像硬件上进行了验证(出于监管目的)。

医院和放射科诊所(行业利益相关者) (Hospitals and Radiology Clinics (Industry Stakeholders))

Must be able to support the infrastructure for any algorithm deployed (cloud resources / GPUs / other hardware)

必须能够支持所部署的任何算法的基础架构(云资源/ GPU /其他硬件)

Interest in AI is usually tied to reducing costs over time.

对AI的兴趣通常与降低成本有关。

监管机构(监管利益相关者) (Regulatory Bodies (Regulatory Stakeholders))

Their main concern is for the safety of patients. That means very tight regulations.

他们主要关心的是患者的安全。 这意味着非常严格的法规。 They ensure that the algorithm is labeled, marketed, and used appropriately in a clinical setting.

他们确保在临床环境中正确标记，销售和使用该算法。

Regulatory bodies like the FDA treat AI algorithms for medical imaging like medical devices.

像FDA这样的监管机构将AI算法用于医疗设备等医学成像。

人工智能什么时候最有效？ (When is AI most effective?)

When it makes the job of everyone involved easier and more efficient. The accuracy of your algorithm matters but the way it’s implemented can be just as important.

当它使每个人的工作变得更加轻松和高效时 。 算法的准确性很重要，但是其实现方式可能同样重要。

Algorithms should make critical studies FASTER to interpret.

算法应该使批判性研究更快地进行解释。

It should help reduce burnout and fatigue for clinicians.

它应有助于减少临床医生的倦怠和疲劳。

Make accurate assessments.

进行准确的评估。

Be easy to use by those involved.

易于相关人员使用。

案例范例 (Case Examples)

减少放射科医生的倦怠 (Reducing Radiologist Burnout)

Application: Screening for mammography

应用范围 ：乳腺X线摄影检查

Summary: Millions of mammographies are taken a year in the United States and require 2 radiologists to interpret the results. Over 85% of the mammographies taken end up being completely normal.

简介 ：在美国，每年需要进行数百万例乳房X光检查，并且需要2位放射科医生 来解释结果 。 最终，超过85％的乳腺摄影完全正常。

Problem: Radiologists are wasting hours interpreting normal mammographies.

问题 ：放射科医生浪费大量时间解释正常的乳腺X线摄影。

Potential Solution: Use classification. Depending on the accuracy of the model, algorithms may be used to determine whether a mammogram comes back positive or negative, and then a single radiologist can be used to verify the result.

可能的解决方案 ：使用分类。 根据模型的准确性，可以使用算法来确定乳房X线照片返回的是正值还是负值，然后可以使用一位放射线医师来验证结果。

在急诊室环境中为放射科医生优化工作流程 (Optimizing Workflow for a radiologist in an ER setting)

Application: Flagging brain bleeds

应用 ：标记脑出血

Summary: Radiologists read images in a first-in first-out queue from the PACs system.

摘要 ：放射科医生从PACs系统中以先进先出队列的方式读取图像。

Problem: Brain bleeds are fatal if left untreated and need to be quickly discovered.

问题 ：如果不及时治疗，脑出血是致命的，需要Swift发现。

Potential Solution: A classification algorithm that is run before placing the image into PACs. If it determines that a brain bleed exists, it can be pushed to the front of the queue for a radiologist to verify.

可能的解决方案 ：在将图像放入PAC中之前运行的分类算法。 如果确定存在脑出血，则可以将其推到队列的最前面，以便放射线医生进行验证。

Taking into consideration all these components and having a more holistic understanding of the medical landscape is the best way to ensure that your algorithms can actually be used in the medical setting.

考虑到所有这些组成部分并对医学领域有更全面的了解，是确保您的算法可以在医疗环境中实际使用的最佳方法。

If you enjoyed this piece consider checking out some of my others!

如果您喜欢这个作品，可以考虑看看我的其他一些作品！

翻译自: https://towardsdatascience.com/ai-for-healthcare-an-introduction-f5ae368bc0ef

相关资源：科特迪瓦卫生保健人员进行新生儿复苏的知识和实践