Plusman: Mie University was our very first customer. We heard that even before the 2020 implementation, you had already been considering AI integration. Compared to other institutions, you moved early. What were your expectations and goals?
April 2022
Mathematical Medicine Vol. 2
Interview with a Leading Expert in Medical AI
University Hospitals Leading the Digital Transformation of Community Healthcare
Dr. Hajime Sakuma
Vice President, Mie University
Deputy Director of Mie University Hospital (Clinical Services)
Head of Radiology, Director of Central Radiology and Medical Information Management
Professor, Department of Radiological Sciences, Graduate School of Medicine, Mie University
Dr. Kakuya Kitagawa
Professor, Advanced Diagnostic Imaging, Graduate School of Medicine, Mie University
Establishing an Environment for Implementing AI in Chest CT for All Patients
Drs. Hajime Sakuma and Kakuya Kitagawa of Mie University were among the first to implement the AI tool Plus.Lung.Nodule into clinical practice. With a distinguished background in cardiovascular CT and MRI, they shared their motivations for adopting chest CT-AI, their experiences since implementation, their expectations for new AI tools, system management, collaboration in community healthcare, and the Hoku-sei Satellite Project.
Expectations for Plus.Lung.Nodule
Dr. Sakuma: Most institutions recognized that AI could be valuable for diagnostic imaging, especially in radiology. But a few years ago, AI was mostly treated as an experimental research tool. In truth, AI’s true value is only understood through actual use.At the 2019 RSNA (Radiological Society of North America) conference, I learned about Plusman’s chest CT AI (Plus.Lung.Nodule, approval number 301AGBZX00004000). I contacted your team through Dr. Nagata, and we started discussions.
I felt it was essential to create an environment where AI could be applied to every patient who undergoes chest CT, regardless of the original purpose. For instance, even if the scan was done for the abdomen but included part of the lungs, we wanted to ensure that AI analysis could still be used. So, we built a system within Mie University Hospital that sends any CT image containing lung fields to the AI analysis server.With support from Dr. Nakayama of Ritsumeikan University and Mr. Nakako of ENTORRES, we developed a system where any CT scan done at the hospital—if it includes lung fields—is automatically sent from PACS to Plusman’s AI server. This system was also designed to retrieve and pre-analyze prior CT scans from up to a year ago. Plusman’s system was highly flexible and allowed us to launch and run it in a short time.
Note: Plus.Lung.Nodule detects lung fields from CT images and can analyze any image containing lung fields, even partially.
The second point is that university hospitals are training grounds for young physicians who are not yet certified specialists. Although awareness of AI has grown among young doctors in recent years, back then AI still seemed intimidating and complex. I wanted to create an environment where they could engage with AI naturally—without excessive expectations or unnecessary skepticism—during their training.
Another reason this was possible is because, in addition to my roles as Director of Radiology and Central Radiology, I was also Director of the Medical Information Management Department. That allowed us to flexibly build this kind of hospital-wide system. I’ve also worked closely with Mr. Nakako at ENTORRES over the years, as he’s been part of the radiology department team. The fact that this was a collaboration between a local company and Mie University was key to the fast rollout.
Plusman: With those expectations and goals in mind, was the implementation of Plus.Lung.Nodule successful in achieving them?
Dr. Sakuma: As soon as we confirmed that all CT images containing lung fields were successfully sent and that AI analysis results appeared in PACS, I felt one of our major goals was achieved. I believe that if a tool is genuinely useful, people will start using it naturally—without me having to push them.
Plusman: Why do you insist on using AI to analyze all cases?
Dr. Sakuma: The reason we perform AI analysis on all cases before sending them to the PACS is simple: if introducing AI means it takes longer to interpret images, and it means people have to finish work and get home late, there’s no point. If the doctor reading the images wanted to see the results of the AI analysis, we needed an environment where they could do so immediately with the push of a button, otherwise we wouldn’t have bothered to introduce it.
Plusman: After actually using it, did it meet your expectations? Or not?
Kitagawa: Speaking specifically about Plus.Lung.Nodule, I felt that it was very sensitive and rarely missed lesions. On the other hand, of course, there are some parts that it picks up too much, so I feel that it is necessary to think carefully about how to use it and find a compromise in clinical practice.
It is not necessary to use it on people who are known to have lung lesions, but I think it is quite good to use it in cases where you think, “I think there is probably no disease, but I would like to confirm at the end.” I think the same thing can probably be said about various AIs. For medicines, the package insert or guidelines say, “It would be good to use this when this patient has these symptoms and the test results are like this,” but the package insert for AI software says what it does, but not when it is effective to use it. I feel that it is necessary to find good ways to use AI image diagnosis.
Plusman: Recently, in parallel with the medical fee revision documents, the guidelines for the use of AI (management guidelines for the clinical use of radiological imaging diagnostic support software using artificial intelligence technology) issued by the Japan Radiological Society clarified the establishment of a system for safety management. As you just mentioned, the usage of AI should first be properly verified by radiologists, and radiologists should play a kind of leadership role in the hospital to ensure that AI is used correctly. What kind of system has been established at Mie University Hospital?
Dr. Sakuma: In the clinical practice at the affiliated hospital, AI is used not only in the radiology department but also in the endoscopy room. However, there are no specific guidelines for how the hospital should use AI at the moment. The AI for endoscopy may be more necessary in terms of establishing a medical safety system in that it alerts patients during the examination. I don’t think there is a clear answer yet on how AI should be used for image diagnosis in the radiology department. First, I think it is important to start with everyone gaining experience, and as they accumulate experience, to establish a system for the appropriate use of AI and safety management.
From the perspective of the use of deep learning in medicine, for example, this CT image. Young radiologists who read the images may not be very aware of it, but the CT image also uses a GPU to perform massive calculations to reconstruct the image, and when the image is sent to the PACS, analysis by Plus.Lung.Nodule is performed on all slices of all cases. Without us knowing, huge calculations related to deep learning are being performed before the image reaches the doctor who reads it. So, although AI is actually deeply integrated into daily clinical practice in radiology, I think the reality is that we are not aware of it, and I think that this is the correct way to use it.
Plusman: When you say AI has become part of everyday life, does that mean that doctors have accepted it without any sense of discomfort?
Dr. Sakuma: From the perspective of the radiologists who read the images, everyone knows that AI analysis is being carried out on all cases, so anyone who finds it useful can start using it right away. AI technology is also used constantly to reconstruct CT images. I think that to you all, AI has become like water or air.
What about you, Dr. Kitagawa?
Dr. Kitagawa: We recently installed the chest X-ray AI developed by Plusman, which I think is quite easy to use (in April 2022, Mie University introduced the chest X-ray AI Plus.CXR (certification number 301AGBZX00004000)). As I mentioned earlier, there is not much need to use it when it is clear that there is an illness, but when you apply AI when there seems to be nothing at first glance, it feels like a trainee is sitting next to you and asking, “Doctor, what do you think here?” If you say something like, “That’s a blood vessel, so it’s fine,” it is possible that the part the trainee said, “What about here?” turned out to be important, so I think it’s really great to be able to use it in this way, very naturally, and read the AI results side by side.
The Key to Reducing Effort and Shortening Work Hours: Determining Whether the Nodules Are Increasing from Recent Images Alone
The Relationship Between Work Style Reform and AI
Dr. Sakuma: Work style reform is becoming very important in the medical field. University hospitals will be facing a critical period of work style reform in the next few years. I think the upper limit on work hours due to the work style reform will have such an impact that it will significantly change the way medical treatment is done. AI is effective in preventing oversights as a diagnostic aid, but when you try to use AI, it often requires more mouse operations and takes longer to read images. If doctors have to spend extra time and effort using AI, it will have a negative impact on patient treatment from a broader perspective.
Pilots and truck drivers have various restrictions, such as not being able to work more than a certain number of hours in a row and having to take a minimum of a certain number of hours off before working the next day. Radiologists are people, so it will become more difficult to work more than ten hours a day and look at tens of thousands of images. AI support is important, but I think that AI systems that make doctors go home late will eventually become unusable for medical treatment.
Dr. Kitagawa: Inevitably, the amount of information we have to look at will increase, so what we used to have to review two or three times will no longer need to be done three times thanks to AI, and we can reduce the time required to do so. To successfully introduce it, we don’t just need to install it, but we need to work it together with the viewer to make it easier to use, and we need to think about when it can actually be used to shorten work hours and image reading time. I think it would be great to be able to announce to everyone how to use it. We’re not there yet, though.
Dr. Sakuma: I think the key to reducing labor and shortening work hours is to be able to see at a glance whether the nodule is growing or not on the latest image. At present, the system has a function to automatically select past images and use AI to measure the size of the nodule, but the function to display whether the nodule has grown or not on the latest image has not been realized. I think this will be a very important point from the perspective of work style reform.
Dr. Kitagawa: Yes, that’s right. We are constantly comparing images with past ones, so it would be very helpful if we could be told “this is what is increasing.” With or without AI, I think comparisons would still be made by human eyes. I feel that using AI would make things even more efficient.
Dr. Sakuma: Personally, I think that if this kind of function were available, a lot of facilities would introduce AI.
For Facilities Considering AI Implementation
Plusman: You mentioned earlier that AI is like air. We think that rather than deciding how to use AI and then coming down with a plan, we should try it out first and then think about it. I think that some hospitals are close to being able to process and see what you want to see right now, but I think that Mie University Hospital is probably the only hospital that has a fully automatic AI system that pre-fetches and analyzes images from about a year ago. So, is there anything you can say to hospitals that are considering introducing AI in the future?
Was there any major confusion the day after it was installed, or within a week or so? Or have there been any accidents or disruptions to medical treatment since it was installed until now?
Dr. Sakuma: We didn’t tell everyone to use it right away, but rather introduced it to those who were interested and started using it.
Dr. Kitagawa: It wasn’t open to other departments, so it was just something within the radiology department and there weren’t any particular problems.
Dr. Sakuma: University hospitals do not use AI for screening in health checkups, and since many patients have many findings to begin with, we spend a considerable amount of time reading each case. In that sense, the impact of AI-based lung nodule detection may be relatively limited in university hospitals compared to general hospitals and health checkup facilities. In general hospitals, there may be people who use AI more widely.
Dr. Kitagawa: There is an overall shortage of radiologists in most places, so there are high hopes for AI, but if it is used incorrectly, it could cause confusion on-site.
Expecting the Emergence of AI Beyond Simple Detection
Dr. Sakuma: At Mie University Hospital, there are doctors in cardiovascular surgery and pediatric cardiology who are interested in AI, and they have written research papers on AI. Therefore, there is a high level of interest in AI at our hospital, but since the doctors are specialists in congenital heart disease, they use AI to predict the Qp/Qs index, which indicates whether pulmonary blood flow is increasing, from simple chest radiographs. Therefore, in terms of expectations for future AI, it may be that the current AI related to radiology is too limited in its purpose to prevent the overlooking of pulmonary nodules and ground-glass opacities. For example, in simple chest radiographs of preoperative patients at university hospitals, it is important to check whether there is cardiac enlargement or abnormalities in the large vascular system, and there are probably few doctors at university hospitals who are asked to perform simple chest radiographs for the purpose of detecting pulmonary nodules.
Dr. Kitagawa: It doesn’t happen that often. When they really need to find a nodule, they order a CT scan.
Dr. Sakuma: At Mie University Hospital, we have overflowing echocardiography slots, so we do not perform preoperative echocardiography on non-cardiac surgery patients unless they have symptoms such as shortness of breath. Chest plain radiographs are often used as preoperative tests to check for cardiovascular abnormalities such as heart failure or incidental lung abnormalities. Assuming that AI-assisted diagnosis of chest plain radiographs was introduced throughout the hospital, even if detailed notes such as “This AI system only looks at nodules” were added, I don’t think busy doctors in each department would read them. Therefore, there is a possibility that doctors in other departments will misunderstand and think, “There’s nothing wrong with the AI.” That’s what I’m afraid of.
Utilization of AI Plus.CXR in Chest X-ray
Dr. Sakuma: I think AI analysis of plain chest radiographs is suitable for general hospitals and health checkup facilities, but it is not easy to try out AI for plain chest radiographs at a university hospital, considering the backgrounds of the patients.
Dr. Kitagawa: There are many different cases, so it may be good to try it out, but it’s difficult to prove its effectiveness at a university hospital. University hospital cases are not suitable for purposes such as screening for nodules.
Dr. Sakuma: For purposes such as finding lung metastases, CT scans are requested from the start. Rather, we need AI that can respond to such purposes as checking for cardiac enlargement, pleural effusion, and abnormalities in the cardiovascular system.
Dr. Kitagawa: I have the impression that it is surprisingly able to detect fibrosis caused by interstitial pneumonia.
Management of AI by Radiologists
Plusman: I see. That sounds very persuasive, and while it does seem like you want to try it out first, it’s still difficult to get the whole hospital to use it all at once. So is it necessary to use the AI to check what it is trying to detect and whether there is any possibility that it may have overlooked something, and then provide guidance?
Dr. Sakuma: Doctors in each department at university hospitals are extremely busy, so it’s rare for someone in another department to voluntarily try using AI on their own. Explaining in detail how to use the AI, what kinds of diseases it’s useful for, and what kinds it’s not suitable for, to the hundreds of doctors across the hospital—I don’t think that’s realistic.
Plusman: So, is that why there’s a need for radiologists to manage it?
Dr. Sakuma: Radiologists are already fully occupied with interpreting CT, MRI, and nuclear medicine scans, and there are few facilities where radiologists can read all of the plain X-rays in the hospital. With the introduction of AI-assisted interpretation for plain X-rays, the radiology department has started managing AI usage by other departments. But if the radiology department says, “We don’t read plain X-rays,” that’s just not going to work, is it?
Mie University Hospital eases local medical institutions' concerns about AI and provides better medical care to patients
Regional Medical Cooperation and the Hokusei Satellite
Plusman: What kind of project is Plusman Hokusei Satellite?
Dr. Sakuma: Hokusei Satellite Medical DX is a project that will begin this April at Mie University. Contributing to the local community is extremely important for a local national university. Mie University’s School of Medicine dispatches doctors throughout Mie Prefecture, fulfilling an important mission as a university, but in terms of directly examining local residents, the university hospital is somewhat less involved. Mie University’s Hokusei Satellite Medical DX project will be based at Kuwana City General Medical Center and will promote medical DX in the northern part of Mie Prefecture.
The importance of online medical care, home medical care, and PHR in regional medical care is rapidly increasing, but even in the field of radiological diagnosis, various medical support software using AI is now available, and I have heard that various companies are selling to small and medium-sized hospitals in Mie Prefecture. At the same time, I have heard from the directors of each hospital that “I don’t know which AI diagnostic support software to choose” and “I don’t know how to use it.” There also seems to be some concern that if their hospital signs a contract with a company’s AI diagnostic support service, they will not be able to switch to another company.
Mie University Hospital has been using AI on all chest CT images for the past two years, and currently has AI from multiple companies in operation. The idea is to evaluate the AI at the Mie University Department of Radiology and Kuwana City General Medical Center, and provide the AI functions that are deemed acceptable to hospitals in the northern part of Mie Prefecture as part of regional medical DX on a cloud-based server installed at Mie University Hospital. In Mie Prefecture, radiologists at Mie University Hospital and related hospitals perform remote image diagnosis from their homes for many hospitals in the prefecture, forming a regional network for image diagnosis. We believe that it is important to provide better medical care to the people of the prefecture by providing quality assurance for AI and connections to image diagnostic equipment as part of regional medical DX and providing AI diagnostic support to local residents and medical institutions in a safe and low-cost manner.
Dr. Sakuma: AI is constantly evolving, so it’s different from CT or MR equipment, which you use for 10 years after you install it. AI functions change rapidly every few years, so I think it’s important for university hospitals to choose AI as the times progress and provide high-quality services within the prefecture as part of regional medical DX.
Plusman: Thank you. I think Mie University is highly relied upon in terms of regional medical cooperation, and I understand that there are very strong ties between Mie Prefecture, Mie University, and each region.
Dr. Sakuma: Mie University affiliated hospitals are located all over Mie Prefecture. Mie University regularly holds meetings with the directors of related hospitals in the prefecture, and I think it is a good place to cooperate with the local community on AI.
Plusman: Local hospitals rely on Mie University, don’t they?
Dr. Sakuma: I think the relationship between Mie University Hospital and its affiliated hospitals is good overall.
Plusman: I thought that was a good example of it working. Are there any tips for building those kinds of relationships?
Dr. Sakuma: Mie University has a history of sending doctors throughout the prefecture. Since they work in cooperation with affiliated hospitals, I think they will be able to do AI as well.
Utilization of Plusman AI at the Hokusei Satellite
Plusman: Recently, we introduced AI for Plain chest radiography (Plus.CXR). Now that you have introduced two types of AI, one for simple chest radiography and one for CT scans, how do you plan to use them?
Dr. Kitagawa: In terms of its use in the Hokuriku Satellite DX Project mentioned earlier, many CT scans and chest X-rays are taken for medical checkups, but there is also a shortage of people to read these.There are high hopes for AI as an auxiliary diagnostic tool, but as Dr. Sakuma said earlier, it is difficult for each hospital to decide on its own to introduce AI, so we are considering having the data uploaded to the Mie University cloud, where we will analyze it and return it to them. We would like to consider various options, including how to best return the results.
Plusman: It is mandatory by law to read health screening twice, but do you think introducing AI would be of any help?
Dr. Sakuma: I don’t think the double-reading rule will change even with the introduction of AI. In terms of education, it might be a good idea to train young teachers to respond to the areas pointed out by AI by saying things like, “This is a normal structure,” or “This is a shadow like this, so it’s not a disease.”
Dr. Kitagawa: That is certainly true.
Dr. Sakuma: I think that if you’re a radiologist you can immediately tell, “This shadow is like this,” but that’s not necessarily the case for all doctors in other departments.
Dr. Kitagawa: The overwhelming majority of areas pointed out by AI in chest X-rays are normal structures. If you’re a radiologist, you can just keep saying, “That’s a normal structure,” but if the reader is insufficiently trained or anxious, there is a good chance that overdetection will occur. If used well, I think it can be expected to speed up the process and reduce oversights.
Dr. Hajime Sakuma
Vice President, Mie University
Deputy Director of Mie University Hospital (Clinical Services)
Head of Radiology, Director of Central Radiology and
Medical Information Management
Professor, Department of Radiological Sciences,
Graduate School of Medicine, Mie University
2022 – Present: Vice President, Mie University
2018 – Present: Vice Director (Clinical Services), Mie University Hospital
2012 – Present: Professor, Department of Radiology, Graduate School of Medicine, Mie University
2009 – Present: Chief, Department of Diagnostic Radiology, Mie University Hospital
2009 – Present: Director, Central Radiology Department, Mie University Hospital
1998: Associate Professor, Department of Radiology, Mie University Hospital
1996: Lecturer, Central Radiology Department, Mie University Hospital
1991 – 1996: Research Fellow, Department of Radiology, University of California, San Francisco
1985: Graduated from the School of Medicine, Mie University
Dr. Kakuya Kitagawa
Professor, Department of Advanced Diagnostic Imaging,
Graduate School of Medicine, Mie University
2020 – Present: Professor, Department of Advanced Diagnostic Imaging, Graduate School of Medicine, Mie University
2019 – Present: Mercator Fellow of the DFG, Department of Radiology, Charité – Universitätsmedizin Berlin
2019: Associate Professor, Department of Radiology, Mie University Hospital
2015: Lecturer, Central Radiology Department, Mie University Hospital
2010: Assistant Professor, Central Radiology Department, Mie University Hospital
2006 – 2008: Research Fellow, Johns Hopkins University
1997: Graduated from the School of Medicine, Mie University