AI won’t replace radiologists, but it will dramatically change their jobs
AI won’t replace radiologists, but it will dramatically change their jobs
人工智能不会取代放射科医生，但会彻底改变他们的工作方式
In 2016 Geoffrey Hinton, the Nobel-winning “godfather of AI,” predicted that radiologists—the physicians who read X-rays, ultrasounds, and other images to help make medical diagnoses—would find themselves replaced by computers within five years. 2016年，诺贝尔奖得主、“人工智能教父”杰弗里·辛顿（Geoffrey Hinton）曾预言，放射科医生（即那些通过解读X光片、超声波及其他影像来辅助医疗诊断的医生）将在五年内被计算机取代。
Today the field can retort by quoting Mark Twain’s famous quip: The report of my death was an exaggeration. Radiology’s ranks are in fact growing steadily, with the number of practitioners expected to expand by 26 percent or more over the next three decades. 如今，放射学界可以用马克·吐温的名言来回击：“关于我死亡的报道被夸大了。”事实上，放射科医生的队伍正在稳步壮大，预计未来三十年内，从业人员数量将增长26%甚至更多。
But what Hinton may have missed about the dynamics of the job market should not obscure his prescience: He was correct that human physicians now have a silicon-based colleague in the room that matches or exceeds their performance. 但辛顿对就业市场动态的误判，不应掩盖他的先见之明：他是对的，人类医生现在确实拥有了一位硅基同事，其表现足以媲美甚至超越人类。
In fact, radiology is far and away medicine’s hot spot for AI, making it a bellwether for the adoption of expert decision-making systems across healthcare and perhaps in other fields. As of early 2026, about three-quarters of the 1,400 AI-enabled medical devices cleared by the Food and Drug Administration were for radiology. 事实上，放射学是目前医学领域人工智能应用最热门的领域，这使其成为医疗保健乃至其他领域采用专家决策系统的风向标。截至2026年初，美国食品药品监督管理局（FDA）批准的1400种人工智能医疗设备中，约有四分之三用于放射学。
Some make physicians more efficient by drafting reports or alerting them to the images that urgently need attention. But other AI tools have the potential to improve on human performance by identifying abnormalities that may not be visible to the human eye, or interpreting images as well as—and sometimes better than—trained radiologists. 一些工具通过起草报告或提醒医生关注急需处理的影像，提高了医生的工作效率。但另一些人工智能工具则有潜力通过识别肉眼难以察觉的异常，或达到甚至超过受过专业训练的放射科医生的影像解读水平，从而提升人类的表现。
For example, an analysis of 43 clinical trials concluded that AI-assisted colonoscopies reveal more polyps than conventional ones. Improving accuracy is important, because the average human error rates involving diagnostic images are estimated to range from 3 to 5 percent, which translates to about 40 million errors worldwide each year. 例如，一项针对43项临床试验的分析得出结论：人工智能辅助的结肠镜检查比传统检查能发现更多的息肉。提高准确性至关重要，因为据估计，涉及诊断影像的人为错误率平均在3%到5%之间，这意味着全球每年约有4000万次错误。
Yet the solution is not as simple as replacing humans with machines. Ten years after Hinton’s sensational prediction, the big question is not whether humans or AI do statistically better at a given task, but how the technical precision of AI can be combined with the experience and flexibility of humans to improve accuracy for the benefit of patients. 然而，解决方案并非简单地用机器取代人类。在辛顿那番轰动性预言提出十年后，核心问题不再是人类还是人工智能在某项任务上表现得更好，而是如何将人工智能的技术精度与人类的经验和灵活性相结合，以提高准确性，从而造福患者。
Working together
协同工作
Finding the best way to design such a collaborative system is not straightforward. Even if AI is more reliable on average than a skilled radiologist at interpreting certain images, it is still going to make some mistakes that humans would not, says radiologist Curtis Langlotz, director of the Center for Artificial Intelligence in Medicine and Imaging at Stanford University. 设计这种协作系统的最佳方式并不简单。斯坦福大学医学与影像人工智能中心主任、放射科医生柯蒂斯·朗洛茨（Curtis Langlotz）表示，即使人工智能在解读某些影像时平均比熟练的放射科医生更可靠，它仍然会犯一些人类不会犯的错误。
That puts radiologists in the role of evaluating each AI decision—the vast majority of which will be correct—and identifying the rare instances when the algorithm got it wrong. 这使得放射科医生的角色转变为评估人工智能的每一个决策（其中绝大多数是正确的），并识别出算法出错的极少数情况。
“This requires a whole mental rewiring,” says radiologist Paul Yi, section chief of intelligent imaging informatics at St. Jude Children’s Research Hospital in Memphis, Tennessee. Physicians have grown used to overruling computers, but this moment is different. “这需要彻底的思维重塑，”田纳西州孟菲斯市圣裘德儿童研究医院智能影像信息学部门负责人、放射科医生保罗·易（Paul Yi）说。医生们已经习惯了推翻计算机的判断，但现在的情况有所不同。
Electronic medical record alerts triggered by rule-based algorithms have been warning physicians for decades about things like potentially dangerous drug interactions. Physicians typically override those warnings about half the time, Langlotz says. “I’m going to get some advice from the computer and I’m going to have to evaluate that in the context of all the other information I have about that patient and just make the best decision,” he says. 几十年来，由基于规则的算法触发的电子病历警报一直在提醒医生注意潜在的危险药物相互作用等问题。朗洛茨说，医生通常有一半的时间会忽略这些警告。“我会从计算机那里获得一些建议，然后我必须在掌握该患者所有其他信息的背景下对其进行评估，并做出最佳决策，”他说。
In contrast, AI image analysis systems in radiology are usually based on neural networks that can identify tumor subtypes, outline the boundaries of lesions, and perform other diagnostic tasks with high accuracy. Unlike rules-based algorithms, which give answers or prompts that doctors can quickly interpret based on their own medical knowledge, neural networks are referred to as “black box” systems. 相比之下，放射学中的人工智能影像分析系统通常基于神经网络，能够高精度地识别肿瘤亚型、勾勒病变边界并执行其他诊断任务。与基于规则的算法不同（后者提供的答案或提示医生可以根据自己的医学知识快速解读），神经网络被称为“黑箱”系统。
These AI models typically do not reveal how they reached a decision, which makes a radiologist’s job of deciding whether to veto it much more opaque. “Is that really an abnormality, or am I missing something that AI with its subtle mind or whatever is identifying?” says Charles Kahn, editor of Radiology: Artificial Intelligence, a publication of the Radiological Society of North America. “And that is really challenging for us.” 这些人工智能模型通常不会透露它们是如何得出结论的，这使得放射科医生在决定是否否决其判断时，过程变得更加不透明。《北美放射学会》旗下刊物《放射学：人工智能》的编辑查尔斯·卡恩（Charles Kahn）说：“那真的是异常吗？还是说我漏掉了人工智能凭借其敏锐思维所识别出的东西？这对我们来说确实是一个挑战。”
Veto power
否决权
Langlotz gives the example of a theoretical AI tool that can detect 95 percent of the lung nodules on a chest CT, while radiologists are known to detect 90 percent. “Some would say, ‘Oh, the AI is better than the radiologists, therefore we should replace all the radiologists with AI,’” he says. “But I will tell you that the radiologist is going to detect some of that 5 percent that were missed by the machine. And that’s because machine intelligence and human intelligence are different kinds of intelligence.” 朗洛茨举了一个理论上的例子：一种人工智能工具可以检测出胸部CT中95%的肺结节，而放射科医生通常能检测出90%。他说：“有些人会说，‘哦，人工智能比放射科医生强，所以我们应该用人工智能取代所有放射科医生。’但我会告诉你，放射科医生会发现机器漏掉的那5%中的一部分。这是因为机器智能和人类智能是不同类型的智能。”
AI can examine every pixel on an image without getting tired or distracted and compare it to every other image it has ever encountered. In contrast, radiologists’ understanding of disease allows them to interpret images in ways that AI cannot. “How do you make sure that when the AI and the radiologists are naturally thinking something different, that the radiologist accepts every time the AI is right and dismisses every time it’s wrong?” says Nina Kottler, chief medical AI officer of Mosaic Clinical Technologies. “It’s not necessarily easy, but there are ways to do it.” 人工智能可以检查图像上的每一个像素而不会感到疲劳或分心，并将其与它遇到过的每一张图像进行比较。相比之下，放射科医生对疾病的理解使他们能够以人工智能无法企及的方式解读影像。Mosaic Clinical Technologies的首席医疗人工智能官尼娜·科特勒（Nina Kottler）说：“当人工智能和放射科医生的想法自然产生分歧时，你如何确保放射科医生在人工智能正确时每次都接受，在它错误时每次都驳回？这不一定容易，但有办法做到。”
With the goal of creating an ideal AI/human team, radiologists around the world are working to learn how to collaborate with AI so they and their patients can reap its benefits. That collaboration requires dealing with unconscious biases that make a physician either rely on AI too much or dismiss its responses inappropriately. So to work with these systems optimally, radiologists need to appreciate the overall reliability of AI while still being able to spot its mistakes. “It does help to know a little bit about how these systems work so t 为了打造理想的人工智能与人类协作团队，世界各地的放射科医生正在学习如何与人工智能合作，以便他们和患者都能从中受益。这种协作需要处理无意识的偏见，这些偏见可能导致医生过度依赖人工智能，或不恰当地忽视其反馈。因此，为了以最佳方式使用这些系统，放射科医生既需要了解人工智能的整体可靠性，又要具备发现其错误的能力。“了解一点这些系统的工作原理确实有帮助，这样……”