打开AI医学“黑箱” 机器人医生也要“持证”上岗
Opening the "Black Box" of AI in Medicine: Robot Doctors Also Need "Licenses" to Practice
人工智能（AI）在医疗行业的应用一直都受到广泛关注。AI研究如何在遵循临床研究范式的基础上寻求变革创新？如何看待AI在医疗实践中的伦理问题？谁将占据AI医疗的制高点？
The application of artificial intelligence (AI) in the healthcare industry has always received widespread attention. How can AI research seek transformative innovation while following clinical research paradigms? How should we view the ethical issues of AI in medical practice? Who will occupy the commanding heights of AI healthcare?
针对上述问题，8月29日，《新英格兰医学杂志》（NEJM集团）与嘉会医学研究和教育集团在上海举行了一场医学AI研讨会，来自全球顶尖的医学专家和业内领先企业分享了AI在医学应用中的实际应用案例以及未来面临的挑战。
To address these questions, on August 29, the New England Journal of Medicine (NEJM Group) and Jiahui Medical Research and Education Group held a medical AI symposium in Shanghai, where top medical experts from around the world and leading industry companies shared practical application cases of AI in medicine and the challenges ahead.
AI系统的能力如何评估？
How Should the Capabilities of AI Systems Be Evaluated?
NEJM集团编辑德莱森（Jeffrey Drazen）教授从AI医疗研究出版论文入手，他援引一篇发表在2020年4月出版的《新英格兰医学杂志》上的论文。该研究使用人工智能检测视神经乳头水肿的效果，通过计算接收工作特性曲线下面积（AUC）、灵敏度和特异性来评估视盘外观分类的性能，并与神经眼科医生的临床诊断参考标准进行比较，AI在某些方面的工作甚至超过了眼科医生。
NEJM Group editor Professor Jeffrey Drazen started with papers published on AI medical research, citing a study published in the April 2020 issue of the New England Journal of Medicine. The study used artificial intelligence to detect papilledema, evaluating the performance of optic disc appearance classification by calculating the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity, and comparing it with the clinical diagnostic reference standard of neuro-ophthalmologists. In some aspects, AI even outperformed ophthalmologists.
德莱森教授表示，机器学习应该对大量眼底照片进一步研究，这些照片的资源非常宝贵，一方面是因为通常受到隐私保护的限制，另一方面是眼科器械设备的高昂价格限制了资源的普及性。
Professor Drazen stated that machine learning should conduct further research on a large number of fundus photographs, which are valuable resources—on one hand because they are usually restricted by privacy protection, and on the other hand because the high cost of ophthalmic equipment limits the accessibility of these resources.
复旦大学上海医学院眼科学与视觉科学系主任、中国研究型医院学会眼科学与视觉科学专委会主委孙兴怀教授对第一财经记者表示：“医学成像中人工智能算法正在实现爆炸式增长，人工智能在眼底病筛查上具有很大潜力。但AI要在临床上应用好，一定要有一个非常强大的神经网络计算机团队的支持。”
Professor Sun Xinghuai, Director of the Department of Ophthalmology and Visual Science at Fudan University Shanghai Medical College and Chairman of the Ophthalmology and Visual Science Special Committee of the Chinese Research Hospital Association, told Yicai reporters: "Artificial intelligence algorithms in medical imaging are experiencing explosive growth, and AI has great potential in fundus disease screening. However, for AI to be applied effectively in clinical settings, it must be supported by a very strong neural network computer team."
他同时指出，人工智能靠大量的前期病例图片输入进行识别面临的局限性。“因为临床上疾病是千变万化的，以前没有输入过的，它是识别不了的。”孙兴怀表示，“有些疾病不只是表现在眼底，还要结合眼部其他的改变来综合分析判断。人工智能只能给出一个倾向性判断，最终还是要有经验的医生综合分析来决定。”
He also pointed out the limitations of AI recognition based on inputting a large number of early case images: "Because clinical diseases are ever-changing, anything that hasn't been input before cannot be recognized." Sun Xinghuai said, "Some diseases are not only manifested in the fundus but also require comprehensive analysis combined with other ocular changes. AI can only provide a tendency judgment, and ultimately, experienced doctors must make a comprehensive analysis and decision."
医疗涉及到人类的健康，对AI医疗的系统进行评估的重要性不言而喻。德莱森教授表示，未来开发出一套具有针对性的关于AI介入诊断精确性的评估指南很有必要。
Healthcare involves human health, so the importance of evaluating AI medical systems is self-evident. Professor Drazen stated that it is necessary to develop a set of targeted evaluation guidelines for the accuracy of AI-assisted diagnosis in the future.
“在评估AI系统能力的时候，一方面是评估AI的知识和能力，需要满足临床设计的目标；另一方面，AI系统应该是一个可以不断进化和升级迭代的系统，让技术的发展适应人类。”德莱森说道。
"When evaluating the capabilities of AI systems, on one hand, we assess AI's knowledge and abilities, which need to meet clinical design goals; on the other hand, AI systems should be continuously evolving and upgrading systems, allowing technological development to adapt to human needs," said Drazen.
麦考瑞大学教授、澳大利亚健康创新研究所所长Enrico Coiera表示，AI系统大致可以分三类，包括辅助医生的AI系统，比如读心电图；另一种是可以替代医生的AI系统，比如筛查眼底照片；第三类是AI系统将来可以做的医生目前还做不了的事情。
Professor Enrico Coiera from Macquarie University and Director of the Australian Institute of Health Innovation stated that AI systems can be roughly divided into three categories: AI systems that assist doctors, such as reading electrocardiograms; AI systems that can replace doctors, such as screening fundus photographs; and a third category—things that AI systems may do in the future that doctors cannot currently do.
“我们认为评估AI的性能，也应该将人机互动的性能结合起来进行评估，而不仅仅是评估算法，AI系统如何最大程度上帮助人类医生，这是一个更为复杂的问题。”Coiera教授表示。
"We believe that evaluating AI performance should also incorporate the assessment of human-machine interaction performance, not just algorithms. How AI systems can best assist human doctors is a more complex issue," said Professor Coiera.
他还强调，医疗AI应该不仅仅作为一种通用的技术，而是可以作为专业行业的定制化技术，这对于专科而言非常重要。
He also emphasized that medical AI should not just be a general-purpose technology but should be customized for specific professional fields, which is very important for medical specialties.
医疗AI如何获得信任？
How Can Medical AI Gain Trust?
专家普遍认为，AI在医疗领域的广泛应用的前提是要在工业界、科学界、医生以及患者群体中获得信任。有趣的是，多位专家将AI医疗系统和无人驾驶系统进行了类比，无人驾驶上路要先考“驾照”，那么AI医生要上岗也需要“持证”。此外，万一AI系统发生事故，那么相关的责任认定也需要有所规范。
Experts generally believe that the prerequisite for the widespread application of AI in healthcare is gaining trust among industry, the scientific community, doctors, and patient groups. Interestingly, several experts drew an analogy between AI medical systems and autonomous driving systems. Just as autonomous vehicles need to pass a driving test before hitting the road, AI doctors also need to be "licensed" to practice. In addition, if an AI system causes an accident, the relevant liability determination also needs to be regulated.
“就像自动驾驶一样，我认为现在的AI还没有完全成熟。”中科院院士、复旦大学附属中山医院心内科主任葛均波教授表示，“作为临床医生，在手术过程中，我们希望AI能够指导医生进行非常精确的操作。目前看来，这一功能还要通过机器深度学习，提高应对各种复杂状况的应变能力。”
"Just like autonomous driving, I think AI is not yet fully mature," said Professor Ge Junbo, an academician of the Chinese Academy of Sciences and Director of the Department of Cardiology at Zhongshan Hospital affiliated with Fudan University. "As a clinician, during surgery, we hope AI can guide doctors to perform very precise operations. At present, this function still requires machine deep learning to improve adaptability to various complex situations."
今年3月，葛均波在西门子医疗的手术机器人系统途灵的辅助下，完成了中国首例机器辅助冠脉介入手术。在机器辅助手术中，葛均波通过手术室外的操纵杆，隔室操控机器人进行手术，让导丝迅速通过复杂的病变，微调冠脉支架系统的位置并释放。
In March this year, Professor Ge completed China's first robot-assisted coronary intervention surgery with the assistance of Siemens Healthineers' surgical robot system, Corindus. In the robot-assisted surgery, Ge operated the robot through a joystick outside the operating room, remotely controlling the robot to perform surgery, allowing the guidewire to quickly pass through complex lesions, fine-tuning the position of the coronary stent system and deploying it.
葛均波认为，机器人目前的操作还无法达到业内专家的水平。不过他仍然相信，AI系统一定能在慢病随访、疾病筛查、基层医生诊断等领域发挥作用，并且随着未来在感知方面能力等提升，可以最终达到人类顶尖医生的水平。“AI现在达不到预期，并不意味着我们就要停止对它的研究了。”葛均波表示，“未来AI一定会在精准性和标准化方面发挥重要作用。”
Ge Junbo believes that current robot operations cannot yet reach the level of industry experts. However, he still believes that AI systems will certainly play a role in chronic disease follow-up, disease screening, primary care physician diagnosis, and other areas, and with future improvements in perception capabilities, they can ultimately reach the level of top human doctors. "The fact that AI hasn't met expectations now doesn't mean we should stop researching it," said Ge Junbo. "In the future, AI will certainly play an important role in precision and standardization."
西湖大学特聘研究员郭天南表示：“对AI的信任是基于AI所创造的价值，但人们也不应该对AI有过高的期望值，仍应循序渐进，同时加大AI在临床的渗透率。”
Guo Tiannan, a specially appointed researcher at Westlake University, said: "Trust in AI is based on the value created by AI, but people should not have excessively high expectations for AI. We should proceed step by step while increasing AI's penetration in clinical settings."
专家还表示，目前很多基于深度学习的系统仍是一个黑箱，涉及到很多参数，在相关领域数据共享方面的工作进展仍然非常有限。
Experts also stated that many deep learning-based systems are still black boxes, involving many parameters, and progress in data sharing in related fields remains very limited.
AI实验室AI医疗首席科学家姚建华表示：“虽然从公司盈利的角度来看，不太可能把所有的源代码都开源，但是用于AI系统训练的数据是很重要的，我们使用什么样的数据来进行模型的训练和检测，数据是否有偏见，是否符合临床实践的价值，这些应该是适合发布给公众了解的。”
Yao Jianhua, Chief Scientist of AI Healthcare at an AI lab, said: "Although from a corporate profitability perspective, it's unlikely that all source code will be open-sourced, the data used for AI system training is very important. What kind of data we use for model training and testing, whether the data is biased, and whether it meets clinical practice value—these should be suitable for public disclosure."
医渡云首席人工智能科学家闫峻表示：“保证数据的质量非常重要，我们需要的是用于临床研究的结构化和标准化的模型。现在即便是非常好的医疗机构提供的数据，很多也不能直接拿来面向训练，需要自然语言处理技术来识别，识别的过程中需要很多医学逻辑。”
Yan Jun, Chief Artificial Intelligence Scientist at Yidu Cloud, said: "Ensuring data quality is very important. We need structured and standardized models for clinical research. Even data provided by very good healthcare institutions often cannot be directly used for training and requires natural language processing technology for identification, which involves a lot of medical logic in the recognition process."
科大讯飞一位医疗AI负责人强调了医疗AI系统的可解释性。“AI医疗的智能化系统如果提供了服务信息，那么医生需要获得解释；如果AI系统作为一个辅助诊疗系统，那么本身也需要可解释性。”这位负责人表示，“在样本的可解释性技术路线之上，我们再对数据做深度分析，寻找输入和结果是否强相关的，最后还需要结构化的医学诊疗知识。”
A medical AI leader from iFlytek emphasized the interpretability of medical AI systems. "If an intelligent AI medical system provides service information, doctors need explanations; if the AI system serves as an auxiliary diagnostic system, it itself also needs interpretability," said this leader. "On top of the technical path of sample interpretability, we conduct in-depth analysis of the data to find whether inputs and results are strongly correlated, and finally, we need structured medical diagnosis and treatment knowledge."
赞助本站
Sponsored by this site
