内容来源：https://blog.google/innovation-and-ai/models-and-research/google-research/blind-sweep-ultrasounds-ai/

内容总结：

一项最新研究显示，人工智能有望帮助资源匮乏地区的孕妇获得关键产检信息。谷歌研究人员与西北大学及Jacaranda Health合作，在肯尼亚内罗毕和芝加哥对2000名孕妇开展研究，训练基层医护人员执行简单的“盲扫”超声波检查，并利用机器学习模型分析影像。结果显示，该AI模型在检测孕龄和胎位方面的准确度可媲美专业超声技师。

研究表明，全球约三分之二的人口无法便捷获得超声波和X光等诊断影像服务，且超声技师严重短缺。传统超声设备体积大、价格高，在偏远地区维修困难且需要稳定电力。手持超声设备虽已改善可及性，但培养一名超声技师仍需两年密集培训。

该研究采用的“盲扫”方法让从未做过超声的医护人员在八小时内即可学会操作。AI模型可分析盲扫视频，判断是否需要重新扫描，并估算孕龄和胎位。所有处理均在设备端完成，无需电力或Wi-Fi。

研究人员表示，孕龄估算对产前筛查、治疗和安全分娩至关重要，估算错误可能导致严重后果。该研究证明了AI在孕产妇保健中的可行应用，有望将专家级护理直接交到一线社区卫生工作者手中。研究人员认为，这一技术的应用潜力不仅限于孕龄和胎位检测，还可扩展至更广泛的医疗场景。

中文翻译：

问科学家：研究人员如何用AI帮助孕妇获得超声检查？

超声检查能为女性提供孕期至关重要的信息——它让准父母得以听到心跳、数清手指和脚趾、确认怀了几个宝宝、查看胎位，以及许多其他信息。但并非每一次怀孕都能获得这种水平的照护。全球约三分之二的人无法便捷地获得超声和X光等诊断影像服务。而且，超声技师的数量也不足以满足所有需要超声检查的人。那么，如果你所在的地方做不了超声检查，该怎么办？

谷歌的研究人员与西北大学和Jacarand Health合作开展了一项研究，探索了一种用AI解决这一问题的工具。该研究包括培训医护人员执行简单的“盲扫”超声操作，并使用机器学习模型来解读结果。通过这种方法，他们能够像训练有素的超声技师一样准确地检测出重要的产前信息，如孕龄和胎位，从而为资源匮乏地区的女性提供更好的信息，让她们在人生中最重大的事件之一到来之前做好更充分的准备。

我们采访了谷歌员工安吉丽卡·威利斯和尼科尔·杨-林博士，进一步了解这项研究——以及它对未来全球超声可及性可能意味着什么。

你在谷歌做什么工作？

安吉丽卡：我是一名软件工程师和AI研究员。我一直热衷于将AI用于社会公益和医疗公平。在孕产妇照护领域，技术进步的价值取决于它所能影响的生命。我的动力是打造真正能触达最需要它的人群的技术。目前，我的重点是让医疗保健在全球范围内更加可及。

尼科尔：我是谷歌内部的妇产科医生。我其实是从公共卫生领域起步的，研究产后出血以降低孕产妇死亡率。这激励我成为一名医生，但我很快意识到，要解决全球医疗可及性不平等的问题，我们需要技术所能提供的规模化能力。我现在在谷歌从事一系列女性健康和消费者健康项目，同时每周仍继续照护患者。

为什么超声检查如此重要？

尼科尔：作为妇产科医生，我们会用超声来回答各种各样的问题。宝宝发育是否正常？有没有异常？我们尤其依赖早期超声来估算孕龄：怀孕多少周了，预产期是什么时候？筛查、治疗、管理以及为安全分娩做准备——这一切都取决于由孕龄确定的时间线。

很多人月经不规律，或者不记录月经周期，所以我们不能总是依赖月经史来估算孕龄。估错可能会带来严重后果。假设有人因为并发症需要提前分娩。你以为分娩的是37周的宝宝，但也许实际上分娩的是34周的宝宝，肺部发育还不够成熟。这些宝宝出生后需要完全不同级别的照护。当我们能够准确确定宝宝的孕龄时，我们就能更好地为宝宝和母亲提供出生后所需的支持。

是什么让人们无法在孕期获得超声检查？

尼科尔：传统超声设备体积大、价格昂贵，构成了获取障碍。当它们在偏远诊所出现故障时，往往没有人能修，而且即便能正常工作，也需要稳定的电力供应。随着手持式超声设备的最新发展，我们看到了让更多女性获得超声检查的更大可能性，尤其是在资源匮乏的环境中。手持设备更小、更便宜，而且很多是电池供电的，因此便于携带、容易获取。

安吉丽卡：但即使便携设备技术取得了进步，成为一名超声技师仍然需要两年高强度的实操培训，这导致全球范围内能够执行产前超声检查的超声技师严重短缺。

请介绍一下你们的研究。你们在解决这些问题上有什么不同做法？

安吉丽卡：我们与Jacaranda Health和西北医学合作，在肯尼亚内罗毕研究了1000名母亲，在芝加哥又研究了1000名。我们发现，我们的机器学习模型能够像训练有素的超声技师一样准确地检测孕龄和胎位。

在传统超声检查中，超声技师必须精确地移动探头才能获得特定测量值。这真的很难做对。我们的研究采用的是“盲扫”超声。操作者按照预先设定的模式在腹部扫动探头。从未做过超声的医护人员可以在八小时内学会操作。

这些盲扫操作者采集的视频帮助我们获得所需的胎儿影像，我们构建了一个AI模型来分析这些视频。它可以告诉操作者是否需要重新扫查。它还能估算孕龄并判断胎位——宝宝处于什么位置。至关重要的是，所有处理都在设备上完成，因此不需要电力供应，甚至不需要Wi-Fi，就能为操作者和母亲提供准备分娩所需的洞察信息。

这对全球孕产妇医疗保健可能意味着什么？

安吉丽卡：我们将这个项目视为一个展示孕产妇医疗保健领域可能性的机会。这项研究证明，我们可以拓宽获得专家级照护的渠道，将其直接交到一线社区卫生工作者手中。

尼科尔：我们研究的目标是展示扩大可挽救生命的超声照护的可及性将如何影响社区，我们正在一步一步地实现这一目标。通过这项研究，我们在证明这是AI在孕产妇医疗保健中一个可行应用方面取得了巨大进展，因为我们可以在不同的患者群体和医疗环境中使用它。但为什么要止步于孕龄和胎位——甚至止步于孕产妇医疗保健呢？超声可以用来做很多事情，比如查看某人在事故或受伤后出了多少血。AI在改善孕产妇健康和整体医疗保健方面有着巨大的潜力。

英文来源：

Ask a Scientist: How are researchers using AI to help pregnant women access ultrasounds?
Ultrasounds give women crucial information during pregnancy — they allow expecting parents to hear a heartbeat, count fingers and toes, check how many babies there are, see how they’re positioned, and so much more. But not every pregnancy gets this level of care. About two-thirds of people across the globe lack ready access to diagnostic imaging services, like ultrasounds and X-rays. And there aren’t enough sonographers to perform ultrasounds for everyone who needs them. So what do you do if you can’t get an ultrasound where you live?
Google researchers explored an AI tool for this problem through a research study with Northwestern and Jacarand Health. The research involved training healthcare workers to perform simple “blind sweep” ultrasounds and used machine learning models to interpret the results. With this approach, they were able to detect important prenatal information, like gestational age and fetal presentation, as accurately as a trained sonographer, giving women in under-resourced areas better information about what to expect ahead of one of the most significant events of their lives.
We sat down with Googlers Angelica Willis and Dr. Nichole Young-Lin to learn more about this research — and what it could mean for the potential future of ultrasound access worldwide.
What do you do at Google?
Angelica: I’m a software engineer and AI researcher. I’ve always been passionate about using AI for social good and healthcare equity. When we look at maternal care, technical advancements are only as valuable as the lives they impact. My motivation is building technology that actually reaches the people who need it most. And right now, my focus is making healthcare more accessible worldwide.
Nichole: I’m the in-house obstetrician-gynecologist (OB/GYN) at Google. I actually got my start in public health researching postpartum hemorrhage to reduce maternal mortality. That inspired me to become a doctor, but I soon realized that to address global inequities in medical access, we need the scale that technology can offer. I now work on a range of women’s health and consumer health projects at Google,a week while continuing to care for patients.
Why are ultrasounds so important?
Nichole: As an OB/GYN, we turn to ultrasounds for all sorts of questions. Is the baby growing appropriately? Are there anomalies? And we rely on early ultrasounds, in particular, to estimate gestational age: How far along is the pregnancy, and when is the baby due? Screenings, treatment, management, and being able to prepare for a safe delivery — that all depends on a timeline determined by gestational age.
Many people have irregular periods or don’t track their periods, so we can’t always rely on period history to estimate gestational age. And there can be serious ramifications for getting it wrong. Let’s say someone needs to deliver early because of a complication. You think you’re delivering a 37-week baby, but maybe you’re actually delivering a 34-week baby with less mature lungs. Those babies need very different levels of care when they’re born. When we’re able to accurately define the baby’s gestational age, we’re much better prepared to provide both the baby and the mother with the support they need after birth.
What’s keeping people from getting ultrasounds during pregnancy?
Nichole: Traditional ultrasound machines are bulky and expensive, creating a barrier to access. When they break in remote clinics, there’s often no one to fix them, and they require stable electricity when they do work. With recent developments in handheld ultrasound devices, we’ve seen greater potential for us to get ultrasounds to more women, especially in low-resource settings. The handheld machines are smaller and cheaper, and many are battery-powered, so they’re portable and easily accessible
Angelica: But even with advancements in portable device technology, it still takes two years of intense hands-on training to become a sonographer, resulting in a global shortage of sonographers who can perform prenatal ultrasounds.
Tell me about your research. What are you doing differently to solve those problems?
Angelica: We partnered with Jacaranda Health and Northwestern Medicine in our study of 1,000 mothers in Nairobi, Kenya, and 1,000 more in Chicago. We found our machine learning model could detect gestational age and fetal presentation as accurately as a trained sonographer.
In a traditional ultrasound, the sonographer has to move the probe around precisely to get specific measurements. And it’s really hard to get right. Our research involved a “blind sweep” ultrasound instead. Operators sweep the probe over the abdomen in a predefined pattern. And healthcare workers who’ve never performed an ultrasound before can learn how to do it in eight hours.
These blind sweep operators gather videos that help us get the imaging we need of the baby, and we built an AI model to analyze those videos. It can tell the operator if they need to redo the sweep. And it can estimate gestational age and indicate fetal presentation — what position the baby’s in. Crucially, all of that processing happens on the device, so it doesn’t require an electricity supply or even Wi-Fi to be able to provide the operator and the mother with the insights they need to prepare for birth.
What could this mean for maternal healthcare worldwide?
Angelica: We saw this project as an opportunity to demonstrate what’s possible in maternal healthcare. The research proves we can broaden access to expert-level care and put it directly into the hands of frontline community health workers.
Nichole: The goal of our research is to show how expanded access to life-saving ultrasound care could impact communities, and we’re getting there one step at a time. Through this research, we’ve made huge strides toward showing that this is a viable application of AI in maternal healthcare because we can use it across different patient populations and healthcare settings. But why stop at gestational age and fetal presentation — or even maternal healthcare? Ultrasound can be used for so many things, like seeing how much someone is bleeding after an accident or an injury. There’s so much potential for AI to help improve maternal health and healthcare across the board.