Early detection is crucial in treating colon cancer. In the U.S., a 1993 report on the National Polyp Study and a 2012 report on the associated cohort study revealed that endoscopic resection of adenomatous polyps, which account for the majority of lesions, reduced the colon cancer incidence rate by 76-90% and the mortality rate by 53%.<\/p>\n
This demonstrates the importance of not overlooking lesions or early-stage cancer during endoscopic examinations. Nevertheless, one report suggests that 24% of cases are actually missed. Another report states that 6% of patients went on to develop colon cancer, even though they had undergone a colonoscopy. Of the root causes, oversight during endoscopic examination accounted for 58%, far exceeding not visiting the hospital (20%), new onset (13%), and insufficient endoscopic resection (9%). Around 2015, there was mounting debate about ways to reduce such oversight.<\/p>\n
Colonoscopy is a procedure in which a physician uses an endoscope to visually examine the inside of the colon on a monitor to detect lesions. Everything is done manually by humans ——<\/span> both moving the endoscope and viewing the images it captures. Accordingly, oversights can occur, depending on the endoscopist’s proficiency level. Other issues include endoscopists being unable to perform a thorough examination due to time constraints, or impaired concentration from fatigue. In some cases, an endoscopist may detect a large polyp but overlook a nearby smaller one, thinking no further lesions are present. Whatever the reason, it should not happen, but because this is a task carried out by humans, the reality is that oversights do occur when various factors overlap. Accordingly, initiatives have emerged to leverage machine assistance. We wondered whether it might be possible to have a machine detect lesions missed by a human and sound an alarm, similar to automatic braking systems in cars. Other ideas included using a special light that would cause lesions to glow, making them easier to find. We considered all kinds of possibilities.<\/p>\n It was just around that time that scientists made a breakthrough in AI supported by deep learning, attracting a great deal of attention. In 2012, the results of a study were published showing that an AI had become capable of recognizing images depicting cats. After being trained on 10 million randomly chosen images, the AI recognized the characteristics within the images and developed the ability to autonomously classify images of cats. Accordingly, we wondered whether it might be possible to train an AI to classify lesions in the colon so that it could support colonoscopy.<\/p>\n We spent a great deal of time and effort training the AI on lesions. We showed it all kinds of images, from typical to atypical cases. It is undoubtedly better to have a large number of training images. However, we questioned how effective it would be to extract 30 frames per second from a 10-second video of a polyp ——<\/span> resulting in 300 images ——<\/span> to train the AI. With this approach, even if the AI was fully capable of identifying similar polyps, its ability to recognize polyps that are not similar would likely be poor. Accordingly, we decided to stop using clips from the same video and instead train the AI on a broader range of lesions using the collection of still images we had amassed over time.<\/p>\n Generally speaking, experts say around 1,000 images are enough to train an AI on a new model. With that number of images of dogs, for example, the AI will be able to properly identify an image of a dog. However, given that we are dealing with matters relating to medical care, including both typical and atypical cases, that number of images is unlikely to be sufficient. We thought that, to guarantee robustness ——<\/span> the ability to maintain stable performance even if unexpected changes occur ——<\/span> the AI needs to be trained on at least 10,000 examples.<\/p>\n Ultimately, we gathered 250,000 original still images of approximately 12,000 lesions. Staff from the National Cancer Center Hospital‘s Endoscopy Division annotated each image and spent well over a year training the AI. These lesion images, taken over roughly five years at the National Cancer Center Japan, came from about 8,000 patients. Only at the National Cancer Center Japan could we have trained our AI on such a large dataset. Other institutions may have had similar numbers of typical cases, but likely fewer atypical ones. And I have heard that few overseas institutions photograph lesions with the same level of detail as in Japan. To explain why detailed, careful imaging is so important, let us take the example of a photograph in which there is a bubble on top of the lesion. If we teach the AI that this is a lesion, we might end up with a situation in which the AI reacts to the bubble and does not react to lesions that do not have bubbles on them. I am glad that we have stored so many clean images of lesions.<\/p>\n This is how we perfected WISE VISION Endoscopy, an AI diagnostic support medical device software application jointly developed by the National Cancer Center Japan. In November 2020, it was approved as a medical device in Japan.<\/p>\n Two monitors are prepared for use in examinations. One monitor shows the video images taken by the endoscopy camera, as usual. The other is the WISE VISION monitor, on which the AI, viewing the same video, carries out analysis and automatic detection in real time. If there is a lesion, the system marks the area with a circle on the monitor and makes a notification sound (Figures 1-3). By feeding back the detected information to the physician in real time, the system enables the endoscopist and the AI to examine the patient together.<\/p>\n Figure 1. <\/span>Endoscopic image analysis to support physicians<\/span><\/strong>The physician operates an endoscope equipped with an ultra-compact digital camera to search for lesions. The video is simultaneously sent to the AI image analysis system. If the AI detects a lesion, it notifies the endoscopist in real time with visual and audible alert, promoting closer observation and reducing the risk of missed lesions.<\/p>\n<\/div>\n Figure 2. <\/span>Scene from an actual procedure using the AI diagnostic imaging system<\/span><\/strong>The left-hand monitor shows the video images taken by the endoscopy camera, as usual. On the right-hand monitor, the AI analyzes the images and notifies the endoscopist of any lesions.<\/p>\n<\/div>\n Figure 3. <\/span>Results of diagnosis based on image analysis by the AI<\/span><\/strong>The system captures the video shot by the endoscopy camera in real time and generates images. Suspected lesions identified by the AI are marked with circles.<\/p>\n<\/div>\n So, what were the results of using AI in endoscopic examinations? Between November 2020 and April 2021, we used AI in colonoscopies of 110 patients at the National Cancer Center Hospital and The Jikei University Hospital. We then compared the results with those of examinations previously performed by the same endoscopists without the aid of AI. The adenoma detection rate (ADR) when AI was not used was 37%. ADR could be described as an indicator of an endoscopist‘s ability; according to a statement issued by the American Society for Gastrointestinal Endoscopy, the target ADR in male patients is 30%. As such, an ADR of 37% meets that standard. When colonoscopies were performed in conjunction with the AI system, the figure rose to 52%, representing an increase of 15 percentage points.<\/p>\n What was especially interesting was the fact that the effect of performing the procedure in conjunction with AI differed according to the endoscopist’s experience (Figure 4). We divided the physicians into those with five years or more of experience in conducting endoscopic examinations and those with less than five years. When we did so, we found that the ADR of endoscopists with at least five years of experience was 41% without AI, rising to 58% when aided by AI. However, in the case of less-experienced (less than five years) endoscopists, the increase in ADR when AI was also used was small, rising from 23% to just 32%. We then investigated the figures for medical specialists certified by the Japan Gastroenterological Endoscopy Society and found an increase from 39% to 58%, demonstrating that ADR improved even when colonoscopies were performed by certified medical specialists.<\/p>\n Figure 4. <\/span>Endoscopists’ adenoma detection rates (ADR) by years of experience of performing endoscopies<\/span><\/strong>There is a marked difference between endoscopists with at least five years of experience and those with less than five years‘ experience, in terms of improvement in ADR when using AI.<\/p>\n<\/div>\n This difference in experience suggests that an endoscopist’s experience and proficiency are required, even when performing colonoscopies in conjunction with AI, because it is still a human who moves the endoscope around the colon. A seasoned endoscopist has a sense for when there might be something in a particular area and will move the endoscope towards it to investigate in more detail. In other words, regardless of whether AI is used, lesions cannot be picked up unless the endoscope is moved into an area where a lesion is suspected. Lesions cannot be detected without maneuvering the endoscope, and it would appear that experience is required to operate an endoscope. I believe the gradual rise in ADR among novices can be attributed to this lack of experience.<\/p>\n An unexpected advantage to emerge from this study was the possibility that patients’ receptiveness to colonoscopies might increase if the procedure was performed in conjunction with AI. When we asked patients for their views on performing the procedure in conjunction with AI before actually entering the examination room, a total of around 50% of patients replied that it was either “Very good” or “Good.” However, when we asked them again after the examination had been completed, the combined total answering either “Very good” or “Good” increased to about 80%. During the examination, we have the patient look at the monitor with us. Conventionally, when an endoscopist found the site of a lesion, they would explain it to the patient, but the patient would not really understand what they were looking at. However, when the AI detects something and marks the site of the lesion with a circle in real time, the endoscopist can explain, “You can see it’s here. Let’s get rid of this,” so it is easy for the patient to understand exactly where it is. When patients were asked whether they would like their next endoscopy to be performed in conjunction with AI, around 80% replied that they would like it to be used next time as well. If performing procedures in conjunction with AI led to an increase in the percentage of people undergoing a colonoscopy, one could say it would be a major advantage.<\/p>\n The Ministry of Health, Labour and Welfare also recognizes the importance of performing procedures in conjunction with AI, and approved an additional charge to cover a lesion detection support program from June 1, 2024, for the purpose of encouraging their widespread use in clinical practice. This means that medical institutions are permitted to add this charge to the patient‘s medical fees if a colon polyp is detected and excised in an AI-aided examination.<\/p>\n We have finally reached the stage at which AI-based systems are beginning to be put into full-scale operation. To conclude this article, I will explain the future prospects for colonoscopies performed in conjunction with AI.<\/p>\n Our near future goal is to train the AI to detect flat and depressed lesions that are difficult for humans to recognize, thereby further improving detection accuracy. These lesions are easy to miss during endoscopies, because their flat or concave surface makes them difficult to distinguish from the surrounding mucosa. We hope to further increase ADR by thoroughly training the AI on images of such lesions. <\/p>\n Research on qualitative diagnosis of lesions is also underway, and efforts are being made to enable AI and endoscopists to work together to reduce diagnostic errors in identifying lesion types. In March 2025, one of the outcomes of this research ——<\/span> an AI that identifies whether a lesion is neoplastic or non-neoplastic ——<\/span> received approval for use as a medical device. With this device, endoscopists can now work seamlessly with AI ——<\/span> from detecting lesions to their pathological assessment.<\/p>\n There are also potential breakthroughs that could be realized in the future. Currently, the AI is trained exclusively on lesions identified by humans. Consequently, the AI cannot detect lesions which humans have missed. AI, whether in endoscopy or other fields, learns from the data provided by human supervisors. But what if AI could learn automatically, without human instruction ——<\/span> able to identify lesions without being explicitly told they were lesions? If AI also recognized lesions that humans had never identified, it could make a major contribution to cancer prevention.<\/p>\n In colonoscopy, it is necessary to examine every nook and cranny of the colon; we also hope AI will evolve to the point where it can autonomously operate the endoscope ——<\/span> like a cleaning robot.<\/p>\n AI-assisted colonoscopy is still in the early stage, but I hope it will significantly reduce the incidence of colon cancer.<\/p>\n The accuracy of endoscopy is steadily increasing. Thanks to high-definition imagingsensors, colonoscopies are able to detect minute lesions, thereby helping to ensure early detection. However, as endoscopists are responsible for image interpretation and diagnosis, missed lesions remain unavoidable. AI-based diagnostic imaging systems aim to address this issue. Trained on 250,000 images of lesions from approximately 12,000 cases, the AI system performed automated analysis in real-time. However, a persistent limitation is that AI system cannot detect lesions that are also missed by humans. Therefore, further evolution of such AI system is demanding.<\/p>\n","protected":false},"author":2,"featured_media":2507,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[14],"tags":[],"class_list":["post-2512","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-medicine"],"acf":{"author":"composition by Takeaki Kikuchi The accuracy of endoscopy is steadily increasing. Thanks to high-definition imaging sensors, colonoscopies are able to detect minute lesions, thereby helping to ensure early detection. However, as endoscopists are responsible for image interpretation and diagnosis, missed lesions remain unavoidable. AI-based diagnostic imaging systems aim to address this issue. Trained on 250,000 images of lesions from approximately 12,000 cases, the AI system performed automated analysis in real time. However, a persistent limitation is that AI system cannot detect lesions that are also missed by humans. Therefore, the further evolution of such AI system is required.<\/p>","person":[{"acf_fc_layout":"personcontent","personimg":2506,"personsholder":"Physician, Endoscopic division, (Concurrent) Department of Genetic Medicine and Services, National Cancer Center Hospital Cat-recognizing AI sparks innovation<\/h2>\n
![\"\"](\"https:\/\/healthist.net\/en\/wp-content\/uploads\/sites\/3\/2025\/09\/292_en_feature01_03_fig01.png\")
<\/small><\/p>\n![\"\"](\"https:\/\/healthist.net\/en\/wp-content\/uploads\/sites\/3\/2025\/08\/292_en_feature01_03_fig02.jpg\")
<\/small><\/p>\n![\"\"](\"https:\/\/healthist.net\/en\/wp-content\/uploads\/sites\/3\/2025\/08\/292_en_feature01_03_fig03.jpg\")
<\/small><\/p>\n![\"\"](\"https:\/\/healthist.net\/en\/wp-content\/uploads\/sites\/3\/2025\/08\/292_en_feature01_03_fig04.png\")
<\/small><\/p>\nAI cannot detect lesions missed by humans<\/h2>\n
illustration by Rokuhisa Chino","intro":"
Visiting Scientist, RIKEN","personname":"Masayoshi Yamada","persondetail":"Graduated from Kanazawa Medical University’s School of Medicine in 2002. After serving as an assistant professor in the Department of Gastroenterology at Kanazawa Medical University, he joined the National Cancer Center Hospital in 2010. In 2014, he was seconded to the Pontifical Catholic University of Chile for three months as an endoscopy instructor. He took up his current post in 2015. In 2021, he received the Japan Cancer Association–Chugai Academy for Advanced Oncology Award (JCA-CHAAO Award)."}],"issue":2493,"custom_css":""},"_links":{"self":[{"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/posts\/2512","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/comments?post=2512"}],"version-history":[{"count":0,"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/posts\/2512\/revisions"}],"acf:post":[{"embeddable":true,"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/issue\/2493"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/media\/2507"}],"wp:attachment":[{"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/media?parent=2512"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/categories?post=2512"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/healthist.net\/en\/wp-json\/wp\/v2\/tags?post=2512"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}