Otsuka: It’s been almost four years since you accepted me as a research collaborator. Juntendo is a global leader in brain MRI research. At that time, did you have any particular expectations for deep learning or AI?
April 2021
Mathematical Medicine Vol. 1
Interview with a Leading Expert in Medical AI
Radiologists Who Don't Get Along with AI Won't Survive
Dr. Shigeki Aoki
Chairman, Japan Radiological Society
Professor, Department of Radiology, Juntendo University Hospital
To Get Along with AI, You Must Get Used to It, Not Just Learn About It
Yujiro Otsuka, representative of Plusman and a deep learning researcher in the Neuron Research Group at the Department of Radiology, Juntendo University School of Medicine, interviewed Professor Shigeki Aoki, who has led AI research, about studies using AI, clinical applications, specialist education, insurance reimbursement, and the future of AI.
AI Use in Research Has Become the Norm
Prof. Aoki: Four years ago, when we were just starting, Geoffrey Hinton’s comments had come out*, and there was a mix of concerns that radiology might disappear and excitement that something new was emerging. So, there were both fears and expectations.
*Note: In 2016, Professor Geoffrey Hinton of the University of Toronto said, “People should stop training radiologists now. It’s just completely obvious within five years, deep learning is going to do better than radiologists because this can be able to hit a lot more experience and it might be ten years but we got plenty of radiologists already.”
Otsuka: As an actuary by profession, I started studying deep learning due to requests from pharmaceutical companies I consulted for. I treated it as one form of mathematical modeling. Over these four years, I’ve conducted numerous studies, published seven papers, and presented at over ten conferences. What does deep learning mean to you now?
Prof. Aoki: Thanks to your contributions, it’s become a regular part of research. At first, we wrote papers solely on AI, but now it’s integrated into standard analysis workflows. AI has become a normal tool—radiology + AI is now more like image analysis + a bit of AI.
Otsuka: Are there any types of research you’d like to pursue using AI going forward?
Prof. Aoki: If it’s research for the sake of AI, I’d like to focus on avoiding missed diagnoses. When interpreting images, confirming the absence of rare conditions is inefficient, so I’d like AI to handle that task.
Sometimes We’re Not Even Aware of What We Miss
Otsuka: On the flip side, what are some of the negative aspects of deep learning?
Prof. Aoki: One concern is misplaced expectations—believing halfway that AI alone is good enough. In Japan, there are unexpectedly many hospitals and regions operating without radiologists. In contrast, in places like Europe, the US, and probably South Korea, even simple imaging like ultrasound often involves radiologists. In Japan, due to rigid hierarchies, clinicians or primary doctors take full responsibility for the patient’s health. So, even if a neurosurgeon has performed surgery, if the patient continues to complain of a headache afterwards, they will definitely go back to the neurosurgeon.
If AI becomes more advanced, the idea that radiologists are unnecessary could spread further. We need to communicate the value of radiologists + AI to show that it leads to better outcomes. In Japan, the discussion sometimes turns into “AI vs. radiologists,” but rather than AI standing alone, we must enhance the value and presence of radiologists through AI.
Also, when clinicians use AI, radiologists should be involved to make it more trustworthy.
Otsuka: Regarding the role of radiologists, for example, there are guidelines for diagnosing lung cancer. They state that a chest X-ray should be taken first, followed by a CT if necessary.
But sometimes lung tumors are incidentally found during kidney scans. These cases are outside the guidelines, and that’s where radiologists are essential—they cover a broad range of specialties. Surveys show that about 30–40% of lung cancers are found by radiologists reviewing scans for other purposes. Given that, why do some still argue radiologists are unnecessary?
Prof. Aoki: People tend to be most familiar with their own field. When reading images for a specific disease, once that goal is met, they’re often satisfied and don’t consider other incidental findings. That’s one factor.
Another factor is how easy it’s become to scan wide areas with CT. In the past, only the lungs were scanned, maybe at 10mm slice thickness, and only the lung window was printed on film. But now we have 1mm slices from chest to abdomen. So we can detect bone metastases, breast cancer, thyroid cancer, kidney cancer—you name it.
Still, many physicians haven’t updated their mental image of CT since they first started using it. For example, they might not realize that bone metastases are clearly visible now—or that breast tissue is being captured.
Otsuka: I’ve studied bone metastases a bit recently, and I’ve found it quite difficult to detect.
Prof. Aoki: Even obvious cases can be missed. I think people have different attitudes about incidental findings.
Otsuka: Detecting abnormalities is a major goal of AI, but it’s far more difficult than detecting a specific target. I now better understand that radiologists are the last line of defense in image diagnostics.
AI Surveillance Using J-MID Data Can Detect the Spread of Abnormal Diseases
Otsuka: Recently, there’s growing support for strengthening the combination of radiologists and AI. For example, Plus.Lung.Nodule*, a pulmonary nodule detection AI developed at Juntendo University, showed a sensitivity of 98.3%.
*Note: Plus.Lung.Nodule was developed by Professor Shigeki Aoki, then Associate Professor Kumamaru, Associate Professor Kazuhiro Suzuki, and Plusman’s Yujiro Otsuka.
Prof. Aoki: Plus.Lung.Nodule is a great help. Its performance is excellent—especially for things like ground-glass opacities (GGO), which other companies’ AI might miss, it detects reliably.
Otsuka: Thank you! I’m glad to hear that—haha.
Plus.Lung.Nodule has a sensitivity of 98.3%, while early reports said the sensitivity of PCR tests for COVID-19 was around 60%, which surprised and worried me. Given that, could AI for diagnostic imaging play a critical role like PCR did during the pandemic?
Prof. Aoki: I think there are two aspects. In early 2020, CT had an important role in Japan where CT is readily available, especially when PCR testing was limited. The Japan Radiological Society even noted on its website that under certain conditions, CT could be used as a substitute (※1). The issue was the lack of PCR, so CT was a temporary alternative. But false positives were a concern—radiologists, and by extension AI, could mistake other interstitial or viral pneumonias for COVID-19. By June, expectations for AI in this context had cooled. Meanwhile, many researchers, supported by grants and investment, began COVID-related projects.
So the Radiological Society thought: what can we do? We built a surveillance system to register suspected COVID-19 cases based on imaging findings. These registrations showed clear trends in the first and second waves. This meant we could potentially monitor outbreaks independently from PCR, offering value from a public health perspective. With CT widely distributed across Japan, this was a uniquely national advantage. But we also thought: it doesn’t have to be done manually.
Seven major university hospitals and the National Center for Global Health and Medicine—Tokyo, Kyoto, Osaka, Kyushu, Okayama, Keio, and Juntendo—are contributing data to the J-MID system (※2). Around 700,000 studies and 200 million images are now housed at Kyushu University. It’s a huge database. It’s too much for humans to review, but AI can flag suspicious COVID-19-like images. This makes large-scale surveillance feasible. I think AI is well suited to that. Machines can keep watching stable cases without fatigue.
By September, Professor Kensaku Mori’s team in Nagoya had developed a COVID-19-specific model. We’re now hoping to implement it for screening and surveillance.
If the AI had been in place early in the pandemic, it could have noticed spikes in unusual viral pneumonia. If the data infrastructure is ready, we could build an AI quickly and monitor outbreaks just by watching for deviations in the patterns.
※1 http://www.radiology.jp/member_info/news_member/20200424_01.html
※2 J-MID is a national-scale imaging database initiative led by the Japan Radiological Society to reform healthcare using big data and AI. It collects large-scale imaging data from nine major facilities across Japan. Prof. Aoki oversees J-MID.
Otsuka: By using clustering techniques that classify based on various features in the lung fields, we can constantly monitor for abnormalities or detect when features that are not usually seen begin to appear.
Professor Aoki: Exactly. If something like that had existed from the beginning of the COVID-19 pandemic, we probably would have noticed something unusual right away. Normally, we don’t perform CT scans for viral pneumonia. Even if the chest shows slight whiteness due to pneumonia, if the symptoms are mild, we usually don’t go as far as taking a CT. But when COVID-19 started spreading, everyone began getting CT scans. So if we had been conducting surveillance at that time, I think we would have noticed the sudden increase immediately. That’s why I believe that collecting data on various diseases in something like J-MID and having AI classify them to some extent would help us notice when the classifications start to shift. I think that kind of application suits AI well. It’s one way to use AI—if we have big data and the capability to analyze it, even using samples could work. Every day, we’re getting 100,000 image data, and if AI analyzes just a sampled portion of that, I believe it would be sufficient for surveillance. Also, what’s more, each image comes with a report. It might even be easier to use the report itself. Have AI analyze the reports and detect unusual terms—words that haven’t been used before—and simply the appearance of such new vocabulary might indicate that some unknown disease is spreading.
AI Should Handle the Overflow from Radiologists’ Workloads
Otsuka: Could you share what you can about insurance reimbursement for AI?
Prof. Aoki: The biggest issue is if it turns into “AI vs. doctors.” That implies AI is replacing physicians in insurance-covered medical care, which would be hard to accept—for both doctors and patients. People would worry about accountability. So having AI replace doctors is a tough sell. Using AI to replace doctors as a starting point for insurance reimbursement is not the right angle.
Instead, the model should be: “AI + doctor is better than doctor alone.” But if we only allow AI use by radiologists, then a circular argument arises—“The radiologist already reviewed the image carefully, so why do we need AI?” That makes it hard to justify reimbursement.So the better case is when AI improves workflow, or increases reading volume in less time. In Japan, where testing volume is high, this could work. Say one radiologist spends 10–20 minutes per patient, but with AI, it takes 5 minutes, allowing more patients to be read—that extra capacity could justify reimbursement.
In Japan, about half of imaging isn’t read at all, so letting AI handle the overflow is totally acceptable. Globally, it’s harder. In the U.S., for example, radiologists get paid well and are expected to write perfect reports. If not, they face legal consequences. So, for AI to enter that workflow, it must help radiologists read more efficiently—and reimbursement would likely only cover that efficiency gain.
In Japan, there are not many cases where insurance reimburses the amount of work that doctors make easier thanks to AI. In the end, rather than making the amount of work that doctors make easier by using AI eligible for insurance reimbursement, I think the idea will be to increase the number of images read and have the AI analyze the amount that exceeds the number that doctors can read. At least at first, insurance will reimburse radiologists for image diagnostic AI. Since Japan does not have enough radiologists, I think it would be fine to introduce it by holding explanatory meetings at the time of introduction.
To Use AI Clinically, We Must Understand What’s Behind the Numbers
Using Imaging AI Without Radiologists is Dangerous !?
Prof. Aoki: A real danger is if other specialties start thinking they don’t need radiologists because they have AI. That should absolutely be avoided—at least in the beginning.
For example, China’s COVID-19 AI had extremely high sensitivity and specificity. But you, Mr. Otsuka, would recognize that’s because in high-prevalence areas with mostly young patients and few confounding imaging findings, most CTs showed COVID. If non-COVID viral pneumonia wasn’t spreading, then saying “everything is COVID” would yield high performance. But people don’t always grasp that nuance—so it’s risky.
The same applies to cancer: In Japan, many MRIs are done for screening or in early stages, so cases tend to be low-stage. In other countries, where MRIs are rare and expensive, they’re used only for patients suspected of having advanced disease. So imported AI trained on those populations performs poorly in Japan—especially on early-stage cases. Radiologists can understand that kind of bias. But if AI is used by non-radiologists without understanding the background, it’s dangerous.
Otsuka: I read a recent U.S. article highlighting the gap between AI’s catalog specs and its real-world performance. How should we evaluate AI to identify what’s actually usable?
Prof. Aoki: Japan and other countries are in very different situations. I think when importing AI, we should test it in Japan—even with just 100 cases.Also, different hospitals have different patient populations and test practices. Unless you’ve rotated across departments like a radiologist, you might not notice the bias. That means even within Japan, hospital-to-hospital variation is likely. So initially, AI should be used by people who can make informed judgments. If a hospital lacks experts and leaves image interpretation to AI, it could cause serious confusion.
Radiologists Should Lead in AI Education
Experience Before Theory
Otsuka: Given the emphasis on “Radiologist + AI,” what efforts do you think are needed to train radiologists who can effectively use AI for diagnostic support?
Prof. Aoki: I haven’t thought too deeply about it, but I believe it’s important to just start using it. That’s why we launched the Artificial Intelligence Study Group within the Japan Radiological Society. We’ve held hands-on sessions several times. We couldn’t do it at last year’s conference, but we’re planning to this year. The idea is, “Don’t just study it—use it.” Also, when vendors develop new AI tools, we attend their info sessions. I try to promote such opportunities within my circle.
Otsuka: That makes sense. People sometimes ask me, “What’s a good book on deep learning?” or “How should I study it?” But I often say the best way is to actually try it on a real problem.
Prof. Aoki: Exactly. However, you need to have someone knowledgeable nearby. So we try to make sure that at least one person in a department or research group is familiar with AI. That’s why we’ve conducted those hands-on sessions—to create such environments across institutions.
Creating Clinical Guidelines for AI Use
Prof. Aoki: We’re also working on creating guidelines for AI use in clinical practice. This would form the basis for insurance reimbursement. Of course, if we make it too grand and conceptual, it’ll never get done. But we want to at least develop practical clinical usage guidelines soon.
Otsuka: So this is an initiative by the Japan Radiological Society to develop AI usage guidelines?
Prof. Aoki: Yes. When AI is used clinically, we’ll probably start by recommending training through society workshops, understanding metrics like sensitivity and specificity, and requiring that AI be used after reviewing the image oneself. In other words: don’t cut corners.
AI Education Should Be Part of Radiology Certification
Otsuka: Do you think those AI usage guidelines could eventually be tied to board certification? For example, could AI-related education become part of the requirements for becoming or maintaining certification as a radiology specialist?
Prof. Aoki: That’s not in place yet—but I think it should be. We now have a Specialist Certification Organization that oversees 19 clinical fields (including general medicine). Before this, each specialty had wildly different standards—some required training, others only exams, or even just attending seminars.Now, the organization provides structure—e.g., how many cases you need to see or how many lectures to attend. But since this structure is new, it’s currently difficult to add AI to the radiology board requirements. Still, we absolutely should. I really think we should. We should add AI-related sections to the board exam and promote radiology-led AI education to the outside world.
Opening Up J-MID Data for External Use
Otsuka: Regarding J-MID research, it gathers imaging data from seven major universities under the Japan Safe Radiology initiative. The system is designed to provide feedback through data analysis across all stages—from equipment to diagnosis. COVID-19 diagnostic AI was one achievement, and the cloud-based system is expected to make access even easier. Will J-MID be opened up for use by external researchers, companies, or for proposal-based data analysis?
Prof. Aoki: Yes—that’s exactly the direction we’re reforming. The AMED project funding ends this fiscal year. For ethical reasons, users must either belong to one of the seven universities or to the National Institute of Informatics (NII). Even now, data from the seven universities must pass through NII to be accessed, but we’re revising the system so that members of the seven universities can use it more freely.We’re also exploring how to let new universities, research institutes, or companies access the system—provided they pay. So yes, we’re working toward broader access.
The Future of Radiologists and AI
Otsuka: Finally, what are your thoughts on the future of radiologists and AI?
Prof. Aoki: Everyone says it, but it’s true—radiologists who don’t get along with AI won’t survive. Since radiologists are image experts, they’re in the best position to utilize AI effectively. It’s important we harness that advantage to provide better readings and ultimately benefit our patients.
Otsuka: Thank you very much for your time today.
Yujiro Otsuka
Research Fellow, Graduate School of
Medicine, Juntendo University
Representative Partner, Plusman LLC
Consultant, Milliman, Inc.
2006: B.Eng., University of Tokyo
2007: ShinNihon Audit Corporation, Financial Audit Division
2008: Consultant, Milliman, Inc.
2017: Research Fellow, Graduate School of Medicine,Juntendo University
2019: Representative Partner, Plusman LLC
Dr. Shigeki Aoki
Professor, Department of Radiology,
Juntendo University School of Medicine
Chairman, Japan Radiological Society
1984: M.D., University of Tokyo
1995: Associate Professor (Vice Chair), Department of Radiology,
Yamanashi Medical University
2000: Assistant Professor (Associate Professor), Department of Radiology,
Graduate School of Medicine, The University of Tokyo
2008: Professor, Department of Radiology, Graduate School of Medicine,
Juntendo University