Portal de Periódicos da CAPES

Artificial intelligence for early gastric cancer: early promise and the path ahead

2019; Elsevier BV; Volume: 89; Issue: 4 Linguagem: Inglês

10.1016/j.gie.2018.12.019

1097-6779

Yuichi Mori, Tyler M. Berzin, Shin‐ei Kudo,

Metastasis and carcinoma case studies

Artificial intelligence (AI) for GI endoscopy is an important and rapidly growing area of research. Much initial work in AI for endoscopy has focused on detection and optical diagnosis of colon polyps. However, AI has the potential to aid clinical decision making in many other aspects of gastroenterology.1Alagappan M. Brown J.R.G. Mori Y. et al.Artificial intelligence in gastrointestinal endoscopy: the future is almost here.World J Gastrointest Endosc. 2018; 10: 239-249Crossref PubMed Google Scholar In this issue of Gastrointestinal Endoscopy, Zhou and colleagues2Zhu Y. Wang Q.-C. Xu M.-D. et al.Application of convolutional neural network in the diagnosis of the invasion depth of gastric cancer based on conventional endoscopy.Gastrointest Endosc. 2019; 89: 806-815Abstract Full Text Full Text PDF PubMed Scopus (185) Google Scholar explore the potential of AI to address one of the most clinically important issues in the management of early gastric cancers (EGCs): prediction of invasion depth. Effective endoscopic approaches for EGC treatment have developed rapidly. Endoscopic submucosal dissection (ESD) and EMR are now routinely offered to many EGC patients, with surgical options considered only when endoscopic treatment would not likely be curative. The optimal resection approach for an individual patient is determined primarily by predicted depth of invasion, histologic type, lesion size, and the presence or absence of ulceration.3Ono H. Yao K. Fujishiro M. et al.Guidelines for endoscopic submucosal dissection and endoscopic mucosal resection for early gastric cancer.Dig Endosc. 2016; 28: 3-15Crossref PubMed Scopus (344) Google Scholar As a general rule, ESD/EMR should be offered for EGCs with invasion depth limited to the mucosal layer (M) or shallow submucosal layer (SM1 ≤500 μm from the muscularis mucosa), whereas surgery is favored for EGCs invading the deep submucosal layer (SM2 >500 μm invasion). This is because M/SM1 cancers have a very low risk of nodal metastasis, whereas SM2 cancers harbor considerable metastatic potential. However, previous studies have demonstrated that endoscopists achieve only 69% to 85% accuracy4Sano T. Okuyama Y. Kobori O. et al.Early gastric cancer: endoscopic diagnosis of depth of invasion.Dig Dis Sci. 1990; 35: 1340-1344Crossref PubMed Scopus (154) Google Scholar, 5Abe S. Oda I. Shimazu T. et al.Depth-predicting score for differentiated early gastric cancer.Gastric Cancer. 2011; 14: 35-40Crossref PubMed Scopus (72) Google Scholar, 6Choi J. Kim S.G. Im J.P. et al.Comparison of endoscopic ultrasonography and conventional endoscopy for prediction of depth of tumor invasion in early gastric cancer.Endoscopy. 2010; 42: 705-713Crossref PubMed Scopus (167) Google Scholar in predicting invasion depth using endoscopy, EUS, or both. Therefore, accurately predicting the invasion depth of EGCs remains an important clinical challenge. Zhou and colleagues2Zhu Y. Wang Q.-C. Xu M.-D. et al.Application of convolutional neural network in the diagnosis of the invasion depth of gastric cancer based on conventional endoscopy.Gastrointest Endosc. 2019; 89: 806-815Abstract Full Text Full Text PDF PubMed Scopus (185) Google Scholar report a potentially significant advance in the endoscopic assessment of EGC. They developed and validated an AI model that used a deep learning algorithm for determining EGC invasion depth (“M/SM1” vs “SM2 or deeper”). Seven hundred ninety images of gastric cancers were used for machine learning, with a test set of 203 images, the latter being completely independent from the training set. The AI model demonstrated 76% sensitivity and 96% specificity in identifying “SM2 or deeper” cancers and achieved significantly higher performance than endoscopists’ visual inspection. Seven hundred ninety images is a relatively small number in terms of the material for deep learning, but the investigators strengthened the diversity of visual data by using images from 790 different patients (rather than using multiple images from a smaller number of subjects, which is a common approach in this field). This strategy led to excellent performance of the AI model despite the small number of learning images. The very high reported specificity of 96%, if applicable in a real-world clinical setting, would help minimize overdiagnosis of invasion, thereby reducing the likelihood of unnecessary surgery for endoscopically resectable M/SM1 cancers. This type of AI application could also help standardize the assessment of invasion depth by endoscopists, even for patients whose conditions are being evaluated at less-experienced centers. Although the reported results for AI in EGC are promising, we must bear in mind that the present study used an experimental design that has some limitations. In fact, this is the case for many recent studies for AI in endoscopy1Alagappan M. Brown J.R.G. Mori Y. et al.Artificial intelligence in gastrointestinal endoscopy: the future is almost here.World J Gastrointest Endosc. 2018; 10: 239-249Crossref PubMed Google Scholar because early investigation in this field requires laborious collection and classification of images, dividing them into training and test sets, and then assessing and refining the developed model. Such experimental studies of course are indispensable in this early phase of AI research; however, there remains considerable selection bias, as the authors recognized in their article as a limitation. For example, only high-quality images were selected as both learning and test materials. In this case, the developed AI model may demonstrate excellent performance on the test set, but performance may suffer when the software is confronted with the varying image quality and distractors that endoscopists routinely encounter in real life (eg, mucus on the lesion surface, blur, noncentralized lesions, or insufficient air insufflation).7Mori Y. Kudo S.E. Misawa M. et al.Real-time use of artificial intelligence in identification of diminutive polyps during colonoscopy: a prospective study.Ann Intern Med. 2018; 169: 357-366Crossref PubMed Scopus (248) Google Scholar Therefore, before AI technologies like this can be adopted as part of routine endoscopic practice, 2 key steps are required: (1) validation of the model using “real-world” quality images, or better still, unaltered video as a test set, and (2) prospective evaluation with real-time use of the AI model during gastroscopy. This will not only allow more clinically relevant assessment of AI performance but also provide insight on other important real-world considerations, including whether use of AI affects the duration of endoscopic examination, whether AI benefits endoscopists of varying expertise levels, and, most importantly, whether the use of AI achieves a measurable clinical benefit for patients (specifically, more appropriate triage to endoscopic resection or surgery). We expect that to achieve high performance in rigorous clinical studies for EGC, AI systems will need to train on much higher numbers of images, with varying image quality and tumor characteristics. The AI applications for colonoscopy are already several steps ahead on this path. For example, Byrne et al8Byrne M.F. Chapados N. Soudan F. et al.Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model.Gut. 2019; 68: 94-100Crossref PubMed Scopus (329) Google Scholar validated their AI model for polyp diagnosis using unaltered videos containing low-quality image frames as a test set and confirmed that the model provided over 90% accuracy in differentiating between adenomatous and hyperplastic polyps. This level of performance was achieved with a training set of 60,089 image frames. We suspect that the current model proposed by Zhou and colleagues2Zhu Y. Wang Q.-C. Xu M.-D. et al.Application of convolutional neural network in the diagnosis of the invasion depth of gastric cancer based on conventional endoscopy.Gastrointest Endosc. 2019; 89: 806-815Abstract Full Text Full Text PDF PubMed Scopus (185) Google Scholar may require much more real-world training data, possibly numbering in the tens of thousands, to achieve reliably high performance in the clinical setting. However, the workload, time, and resources required to categorize and annotate such a vast number of images is a massive undertaking, and it represents a key hurdle in developing high-performance AI models in clinical medicine.9Cabitza F. Rasoini R. Gensini G.F. Unintended consequences of machine learning in medicine.JAMA. 2017; 318: 517-518Crossref PubMed Scopus (420) Google Scholar Ten years from now, we may look back to the present period as one in which giant leaps were made for AI applications in GI endoscopy. On closer inspection, most giant leaps are actually composed of many smaller steps. The progress reported by Zhou and colleagues2Zhu Y. Wang Q.-C. Xu M.-D. et al.Application of convolutional neural network in the diagnosis of the invasion depth of gastric cancer based on conventional endoscopy.Gastrointest Endosc. 2019; 89: 806-815Abstract Full Text Full Text PDF PubMed Scopus (185) Google Scholar is an important step, but it is not yet a game changer for clinical practice. A great deal of perspiration and collaboration will be required to accumulate the annotated big data that we believe will be a key factor for the giant leaps ahead in the field of AI-assisted endoscopy. Dr Mori and Dr Kudo are recipients of speaking honoraria from Olympus. Dr Berzin is a consultant for Wision AI Ltd. Application of convolutional neural network in the diagnosis of the invasion depth of gastric cancer based on conventional endoscopyGastrointestinal EndoscopyVol. 89Issue 4PreviewAccording to guidelines, endoscopic resection should only be performed for patients whose early gastric cancer invasion depth is within the mucosa or submucosa of the stomach regardless of lymph node involvement. The accurate prediction of invasion depth based on endoscopic images is crucial for screening patients for endoscopic resection. We constructed a convolutional neural network computer-aided detection (CNN-CAD) system based on endoscopic images to determine invasion depth and screen patients for endoscopic resection. Full-Text PDF