Portal de Periódicos da CAPES

Applications of artificial intelligence to endoscopy practice: The view from Japan Digestive Disease Week 2018

2019; Wiley; Volume: 31; Issue: 3 Linguagem: Inglês

10.1111/den.13354

1443-1661

Kazutomo Togashi,

Esophageal Cancer Research and Treatment

With the recent emergence of deep learning,1 artificial intelligence (AI) has stolen the limelight even in the field of gastrointestinal (GI) endoscopy. It is no exaggeration to say that AI is now in its Golden Era. At Digestive Disease Week 2018 held in Washington DC last June, AI-related presentations have increased from three papers (2017) to 30 (2018). AI is one of the most fascinating topics reported by the American Society of Gastrointestinal Endoscopy. At United European Gastroenterology Week 2018 held in Vienna in October 2018, there were approximately 30 AI-related presentations. A similar trend was also observed at Japan Digestive Disease Week (JDDW) 2018 held in Kobe in November 2018. A total of eight papers on the use of AI for endoscopy were presented, most of them associated with colorectal diseases. An overview of these papers is shown in Table 1. Six papers dealt with colonoscopy or stage T1 colon cancer, and the remaining two papers dealt with esophageal cancer and gastric polyps. Notably, six papers trained an AI system using deep learning, suggesting that deep learning will also play a leading role in the field of gastrointestinal endoscopy for a while. The eight papers on the use of AI for endoscopy are outlined here. In the first paper, Ichimasa et al. created a computer algorithm to predict lymph node metastases by machine learning (not deep learning) using text data originating from 590 stage T1 colorectal cancers and examined the diagnostic performance of this approach using another dataset including 110 colorectal cancers from the same institution. Amazingly, the sensitivity for predicting nodal metastases was 100%, specificity 66% and overall accuracy 69%, which is superior to the prediction using the guideline advocated by the Japanese Society for Cancer of the Colon and Rectum. These results are promising, but all data are from one institution. Ichimasa et al. have already started a multicenter validation study using text data from high-volume centers in Japan, to improve the accuracy of the algorithm by analyzing a large quantity of data. If similar or more accurate results are achieved, Ichimasa et al.'s work could dramatically change the management of stage T1 colorectal cancer resected by endoscopy. We really look forward to these results. In the second paper, Misawa et al. created a polyp detecting system from the analysis of 73 colonoscopy videos (including 155 polyps) based on deep learning. In this system, a detected polyp will be noted in real time. They examined the diagnostic performance ex vivo (reading test). Using frame analysis, the sensitivity for polyps was 90%, the specificity 65% and the sensitivity based on the lesions 94%. These excellent results warrant further clinical trials using real colonoscopy. In the third paper, Yamada et al. applied deep learning to 3000 plain endoscopic images of flat or depressed neoplastic lesions and evaluated the ability to identify lesions in another dataset including 751 still images. Sensitivity for polypoid lesions was 97.7% (626/641) and for flat or depressed lesions 97.2% (107/110). These results suggest that AI can detect depressed lesions at the same rate as polypoid lesions, but these results are based only on still images. This observation should be validated using real colonoscopic images. In the fourth paper by Mori et al., they created a polyp-characterization system using ultra-magnified images (CF-H290ECI, Endocytoscopy; Olympus, Tokyo, Japan) using machine learning from a dataset of 61 925 colonoscopic images obtained from a third party which provided data from five hospitals. The Endocytoscope (Olympus) is equipped with a ×520 ultra-magnifying device, providing microvascular and cellular visualization of colorectal polyps after application of narrowband imaging and methylene blue staining. To determine diagnostic performance, they conducted a clinical trial using 436 diminutive (≤5 mm) polyps. In intention-to-treat analysis, surprisingly, the negative predictive value was 95.7%, completely satisfying criteria of the Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI) Statement.2 In other words, this system satisfies the 'resect-and-discard' policy which means both retrieval of resected diminutive polyps and histological assessment can be omitted. In the fifth paper, Kamba et al. created an AI-aided endoscopic diagnostic system with deep learning using 40 000 endoscopic images and performed a reading test for differentiating adenomas from non-adenomas. The sensitivity was 95.2% and negative predictive value 84.6%, suggesting that this system did not satisfy the PIVI criteria. If the application is limited to rectosigmoid polyps, however, the negative predictive value increased to 90%, which was just at the minimum criteria in the PIVI. In the sixth paper, Pu et al. carried out a collaborative study with investigators at Nagoya University, Japan. They created a computer-aided diagnosis system for colon polyps using a modified Sano classification on narrowband imaging, and the average accuracy for this modified Sano classification ranged 82.4–90.9%. Regarding upper gastrointestinal endoscopy, two papers were presented at an early stage and may not have great impact on endoscopy practice mainly because the number of lesions used in the validation process was fairly small (<100) as shown in Table 1. An overwhelming majority of papers relating to AI presented at JDDW 2018 are still on their way to practical applications. The only exception to this was Mori et al.'s paper. They have successfully completed a clinical trial and their AI system may be available together with the ultra-magnification colonoscope Endocytoscopy. Recently, Mori et al. have published these data in a prestigious journal.3 They demonstrated that computer-aided diagnosis for colonoscopy can help endoscopists distinguish adenomas from non-adenomas, not requiring resection. On the first day (1 November 2018) of JDDW 2018, the Japan Gastroenterological Endoscopy Society (JGES) sponsored a concurrent symposium on "Research relating to AI in the Field of GI Endoscopy" in Kobe. The venue was almost standing room only, with 196 endoscopists present. The agenda is shown in Table 2. On behalf of the JGES, Shinji Tanaka, Director of the Central Council for Specialist Systems, made opening remarks. Kiyohito Tanaka, the Special Assistant for President of JGES, gave the first keynote address. He reported that the People's Republic of China is now a leader in AI because of their nationwide campaign, and emphasized that we Japanese gastroenterologists should tackle AI research through an all-Japan effort with governmental support. He started by pointing out that resolution of issues regarding ethics, intellectual property and legislation is the first priority. The second keynote address was given by Dr Tetsuo Sakamaki, of the Japan Agency for Medical Research and Development. He talked about research and development systems, case studies and the expected role of the JGES in AI research. The current progress of AI research was reported by two teams. Dr Shinichi Sato, National Institute of Informatics, and Dr Seichi Uchida, Kyushu University, explained construction of a cloud infrastructure, identification of gastric cancer and automatic classification of endoscopic images based on their site of origin in the digestive tract. Regarding another research team of the JGES, Dr Kiyohito Tanaka explained the Japan Endoscopy Database Project and issues on ethics application. He also reported on-going research projects including typing of gastric cancer, evaluation of inflammatory bowel disease activity, prediction of difficult endoscopic retrograde cholangiopancreatography based on appearance of the duodenal papilla and so forth. After a question and answer session, Dr Hisao Tajiri, President of the JGES, made closing remarks. All of the lectures were informative and educational. Less than one decade has passed since deep learning, which is a key technology, was developed. The progress of AI is very rapid. Endoscopists must learn AI technology and make full use of it. Otherwise, we will not obtain the benefits of AI. The Golden Era of AI has just begun. Authors declares no conflicts of interest for this article. I appreciate Professor Alan K. Lefor (Surgery, Jichi Medical University, Japan) for proofreading.