An artificial intelligence-aided colonoscopy was conducted in yet another clinical trial setting in a hospital in southwestern Chinese city of Chengdu. 

A randomized controlled study – carried out from September 2017 to February 2018 – showed an increase in the detection rate of adenomas by almost 50 percent. A total of 1,058 patients were divided into two control groups: one having gone through a routine colonoscopy while the other underwent an AI-assisted colonoscopy.

Now, the authors of the study plan to refine this system to further validate the effectiveness in screening of precancerous polyps, i.e. adenomas. Adenomas are a sub-type of polyps that potentially have the risk to turn malignant or cancerous.

Doctor Wang Pu, gastroenterologist at the Chengdu-based Sichuan Provincial People's Hospital, said a "tandem" study is underway.

"We first perform a routine colonoscopy, then an AI-assisted one to see how many polyps were missed out in the first procedure and vice versa," Wang told CGTN.

In instances where polyps were found in the first procedure, and which they had taken for biopsy, the residue of polyps were found in the second procedure, Wang said.

"Another rigorous study using a "blind" method – where the endoscopist is not aware that he or she is using the AI system – has been completed. This is to reduce the operational bias by the performing endoscopist."

AI system over human medical assistance?

Li Liangping, chief physician of the hospital's gastroenterology and hepatology department, said plans are also underway to increase the study's sample size.

"We may apply this study to a few other centers, to include doctors at the county level to take on this research. If we can increase their capabilities to increase the detection rate of adenomatous polyps, we can benefit more patients," Li told CGTN.

"In the future, I foresee AI can also be used to enhance the imaging of endoscopy for better accuracy purposes," she added.

Meanwhile, Wang highlighted the possibility of designing a cost-effectiveness analysis of this AI system – compared to that of medical staff assistance.

He said some evidence has shown that with a second observer, a trainee for example, or a nurse, will also help to increase the detection rate of polyps or adenomas.

"But the computer-aided diagnosis system might have less inter-observer variability with a higher fidelity, uniformity and accuracy. We suppose the AI system is more effective than assistants provided by medical staff," Wang noted.

Wang said a follow-up will then be done to assess if the patients who underwent the AI-aided colonoscopy suffer less interval colorectal cancer in maybe five or 10 years. "We'll call them back to receive interval examination."

Liu Xiaogang, chief physician of the hospital's gastroenterology department, noted that AI is only an assistant. "It can replace a lot of our repetitive work, but decision-making is still made by doctors."

How does an AI-aided colonoscopy work?

Research for the real-time automatic polyp detection system began early 2016, Wang said. The algorithm was then validated and published in the Journal of Nature Biomedical Engineering in October 2018.

Chinese doctors from the Sichuan Academy of Medical Sciences and Sichuan Provincial People's Hospital, along with researchers from Israel Deaconess Medical Center and Harvard Medical School jointly came up with this study.

In a nutshell, with the AI system, endoscopists are prompted by blue boxes and alarms of possible polyps while performing colonoscopies. Without the real-time automatic polyp detection system, endoscopists would be relying solely on their eyes to look out for polyps.

Wang said the team collaborated with Shanghai Wision AI to develop the algorithm.

"First of all, we endoscopists collect the images and videos containing a variety of polyps with different morphologists, and we label each polyp existence in the images and each frames of the video clips and we gather these data and send them to the computer scientist, and they train the algorithm used," Wang explained.

"When I insert the colonoscope to the cecum, I activate the system, and I shall check every continuous alarm in the screen provided by the system."

Wang said the system captures the video stream from the colonoscope synchronously, processes each frame and displays the detected polyp location with a blue hollow tracing box directly in the endoscopy monitor.

Significance of an AI-aided colonoscopy

Wang said the value of an AI system is for the easy-to-miss polyps.

"For example, some polyps are partially occluded by faeces or fluids, and some polyps are very flat without a clear boundary, and the color of the polyps is very similar with the background color. So with AI, we can reduce the misdiagnosis of such polyps," he said.

The "miss rates" of polyps during colonoscopic examination varies from six to 27 percent, according to a Gut and Liver study by The Korean Society of Gastroenterology.

Wang brought up the view point of some experts that the study isn't that valuable, as the results of the paper showed that the increase in adenoma detection rate was due to small and diminutive polyps.

"Smaller polyps bear less possibility to progress into cancer. But even though the risk of each polyp is low, the number is huge, the overall expected cancer instance is not negligible. So this system has some significance to the ultimate goal of reducing interval cancer," he said.

Liu, who has been in this line for 25 years, hails this as a breakthrough.

"If my physical strength isn't good, my productivity and accuracy in performing endoscopy in the morning and afternoon would be different. Secondly, my experience 10 years ago and now, is different," Liu said, adding that AI does not have this problem.

"Also, if patients don't prepare well for a scope, even if I want to be on the lookout for polyps, I may not find them," he added.

So it's only by increasing detection rates the increase in large intestinal diseases could be reduced, Liu emphasized.

According to The New England Journal of Medicine, a one-percent increase in adenoma detection rate is associated with a three-percent decrease of interval colorectal cancer.

Limitations of an AI-aided colonoscopy

That said, Liu noted that the AI system presently over-detects on anything that looks like a polyp.

"For example, false positives. The system will identify cells and others as polyps, though instances are a few. It will over-report, but it will not miss out polyps," he said.

"What we can see, they can see. But what we cannot see, they think it is – it goes beyond my range of face recognition ability."

Going forward, Liu said the team is working on further training the AI system, to enable it to better learn and imitate the way humans think, to reduce instances of false alarms.

"We also plan to look into the slight delay of the system, so that it can be used as a long-distance tool. The latency's about 76.80 milliseconds, we humans can't feel it but we hope it will be faster in time."

Screening key to reduce the risk of colorectal cancer

Detecting and removing precancerous polyps is known to be the "gold standard" in preventing colon cancer.

Medical experts say because in most cases these polyps don't trigger any early symptoms, it's hard to know if one has developed them.

This is why it's all the more crucial to get your screening done. If a biopsy confirms they're adenomas which are found, they can then be removed at an earlier stage, thereby eliminating malignant risk.

Wang pointed out that the results of this study may not be generalized to other populations, especially the West.

This is due to the difference in baseline adenoma detection rate for the Western people, which Wang said is higher than that of the Chinese.

"First, in China, the endoscope equipment are older and may not be as advanced as the West. Second, we have different living habits," he pointed.

"Third, there's also the difference in fatigue level. In our hospital, each endoscopist performs 12 to 20 colonoscopies per half day, so the fatigue caused by the heavy workload might be another independent reason for lower adenoma detection rate."