Is artificial intelligence ready to replace expert endoscopists?

Artificial intelligence (AI) is one of the most rapidly evolving new fields of technology over the past decade. We are now in the era of the fourth industrial revolution, which is characterized by the development of robotics, biotechnology, nanotechnology, the internet of things (IoT), big data, and AI, among others, and we have reached a turning point that includes medical services.

“These findings indicate that AI technology has now become available to evaluate all the necessary factors for estimating the risk of lymph node metastasis and determining treatment options, with much better accuracy than experts.”

One key aspect of AI, machine learning, has progressed rapidly because of the technical innovation of deep learning, represented by convolutional neural networks (CNNs). Deep learning is suitable for image recognition and can do this with higher accuracy than humans. Therefore, it is now being increasingly used in daily life, for example for face recognition, self-driving cars, AI chess programs, and so on. It is also very useful in the medical field, especially for radiological image diagnosis, histopathological diagnosis, and endoscopic diagnosis, as these diagnoses are based on pattern recognition by experienced doctors. Among these applications, endoscopic diagnosis seems more challenging than the others because moving images are sometimes used for diagnosis in endoscopy and because endoscope itself is manually controlled by endoscopists, which may affect the quality of images. Despite these difficulties, many studies regarding endoscopic detection and the characterization of gastrointestinal neoplastic lesions have already been published. Among such lesions, gastric cancers present greater challenges for diagnosis owing to the presence of inflammation and undifferentiated-type cancers in the stomach.

Gastric cancer is still very common in Eastern Asia. In Japan, in particular, it was a leading cause of cancer death until a few decades ago; consequently, endoscopic techniques for early detection and treatment became highly developed. As a result, the Japanese guidelines for endoscopic mucosal resection (EMR) and endoscopic submucosal dissection (ESD) for gastric cancer [1] are now regarded as the global standards for endoscopic treatment. According to these guidelines, it is very important to recognize the differentiation, size, depth, and ulceration findings of the tumor when examining cancerous lesions in the stomach. All these factors are essential for estimating the risk of lymph node metastasis in gastric cancer.

Of course, and more importantly, we must detect suspicious lesions during routine esophagogastroduodenoscopy (EGD) while they are still small and flat. Considerable experience is necessary to develop enough skill to detect gastric cancers at an early stage and assess the abovementioned factors. In Japan, images are usually double-checked by an independent expert when EGD is performed as a gastric cancer health checkup. The double-check system is a highly effective way to ensure the quality of the examination and to avoid overlooking cancerous and precancerous lesions, although reviewing a large number of images is usually very tiresome and boring for experts. AI technology is an ideal solution to this problem because endoscopic diagnosis is basically the recognition of a pattern of abnormal mucosa.

The use of AI technology for gastric cancer detection was first reported by Hirasawa et al. in 2018 [2]. They used still images of white-light endoscopy, narrow-band imaging (NBI), and chromoendoscopy, and their overall sensitivity was 92.2 %, although the positive predictive value was 30.6 %. This result seemed reasonable for the first reported use of AI technology to detect relatively small and flat early gastric cancers (EGCs). Soon after this publication, they proved the efficacy of this computer-aided diagnosis (CAD) system for the real-time detection of EGCs using video clips [3]. Prof. Yu’s group in China also reported the usefulness of AI for the detection of EGCs; however, their unique offering was the ability to monitor blind spots during the procedure [4]. They proved the efficacy of their system for real-time monitoring of blind spots in a randomized control study [5]. This system seemed very effective for improving the quality of EGD screening.

Regarding the estimation of the invasion depth of EGCs, another Chinese group has shown the superiority of AI compared with human experts [6]. Additionally, another Japanese group proved the efficacy of AI for this purpose using still images [7]. In addition to these topics, Dr. Ling, a member of Prof. Yu’s group, has now addressed the new challenge of evaluating tumor differentiation and delineating the tumor border using NBI magnification and obtained positive results for both of these purposes [8]. These findings indicate that AI technology has now become available to evaluate all the necessary factors for estimating the risk of lymph node metastasis and determining treatment options, with much better accuracy than experts. For this reason, people might be harboring the illusion that we can automatically obtain accurate information about the risk of lymph node metastasis and suitable recommendations for treatment options with AI.

Unfortunately, however, CAD is not a versatile modality, even though AI has made rapid progress. The thought process of AI is in a kind of black box that seems to work completely differently from experts. In fact, the κ value was very low for the comparison between each expert in Dr. Ling’s study and their CNN1. This suggests that the logic that AI uses to make decisions is completely different from that used by humans, although it can achieve much higher accuracy. Therefore, the final diagnosis should be made by the endoscopist who performed the procedure. More importantly, it is essential to clean the gastric lumen thoroughly before starting an inspection and to take the clearest images possible, as it is almost impossible to obtain a correct answer from AI if the image quality is inadequate. Needless to say, the final decision about treatment options should be made by the responsible doctor with consideration of the patient’s wishes, his or her general condition, underlying diseases, and medication, among other factors.

CAD using AI technology is a wonderful modality for endoscopic diagnosis, but it is not a magic wand. That said, it does already have many merits: (1) shortening the procedure time; (2) reducing the tiresome job of performing double-checks; (3) preventing lesions from being overlooked; and (4) reducing diagnostic disparities, among others. Therefore, if we understand the features of this technology and use it appropriately, it can be of huge help in our daily practice by allowing us, through the detection of lesions at an early stage, to reduce medical costs and provide minimally invasive treatment.

Publication History

Article published online:
22 April 2021

• References

• 1 Ono H, Yao K, Fujishiro M. et al. Guidelines for endoscopic submucosal dissection and endoscopic mucosal resection for early gastric cancer. Dig Endosc 2016; 28: 3-15
• 2 Hirasawa T, Aoyama K, Tanimoto T. et al. Application of artificial intelligence using a convolutional neural network for detecting gastric cancer in endoscopic images. Gastric Cancer 2018; 21: 653-660
• 3 Ishioka M, Hirasawa T, Tada T. Detecting gastric cancer from video images using convolutional neural networks. Dig Endosc 2019; 31: e34-e35
• 4 Wu L, Zhou W, Wan X. et al. A deep neural network improves endoscopic detection of early gastric cancer without blind spots. Endoscopy 2019; 51: 522-531
• 5 Wu L, Zhang J, Zhou W. et al. Randomised controlled trial of WISENSE, a real-time quality improving system for monitoring blind spots during esophagogastroduodenoscopy. Gut 2019; 68: 2161-2169
• 6 Zhu Y, Wang QC, Xu MD. 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: 806-815
• 7 Nagao S, Tsuji Y, Sakaguchi Y. et al. Highly accurate artificial intelligence systems to predict the invasion depth of gastric cancer: efficacy of conventional white-light imaging, nonmagnifying narrow-band imaging, and indigo-carmine dye contrast imaging. Gastrointest Endosc 2020; 92: 866-873.e1
• 8 Ling T, Wu L, Fu Y. et al. A deep learning-based system for identifying differentiation status and delineating the margins of early gastric cancer in magnifying narrow-band imaging endoscopy. Endoscopy 2021; 53: 469-477