Next generation of endoscopy: Harmony with artificial intelligence and robotic-assisted devices

Gulati et al. have reviewed the latest imaging modalities in the field of gastrointestinal (GI) endoscopy as well as the innovations, which are mainly regarding the development of artificial intelligence (AI). The authors have also illustrated hypoxia imaging, robotic-assisted devices to enable automated endoscopy, and eye tracking technology.1 Thus, an intellectual review is beneficial in disseminating the knowledge on how endoscopic imaging is being developed and enabling the readers to foreseeably imagine the future of diagnostic luminal endoscopy.

Artificial intelligence has made significant progress owing to the development of machine learning (ML) algorithms and deep neural networks (DNN). Convolutional neural networks (CNN) is a class of DNN that is highly effective at performing image and video analyses. Advancements in graphical processing unit computing technology have also contributed to higher performance of the deep learning (DL) system at or even exceeding the expert level. CNN is often compared to humans regarding identification, prediction, and detection accuracy; in comparison, DL with CNN can provide a problem-solving method that draws conclusions from the potentially infinite information that cannot be judged by humans quickly. The CNN convolves the images that are reduced in size for further max pooling and threshold-based activation, thus avoiding the risk of overfitting and providing an attractive analytic model for GI endoscopy.2

Computer-aided diagnosis (CAD) in colonoscopy is garnering increased investigation.2-6 AI technology is expected to have two major roles in colonoscopy practice5—automated polyp detection (CADe) and histopathology characterization (CADx). To be maximally effective, CADe should have a high sensitivity for identification with a low rate of false positives and should maintain faster processing speeds to be applicable in real-time during colonoscopy. Recently, in a prospective randomized controlled study, 1058 consecutive patients were randomized to standard white light (WLI) colonoscopy with or without CAD; the polyp detection rate (PDR) and adenoma detection rate (ADR) were significantly higher in the CAD group.3 These effects could be mainly due to a higher rate of detecting diminutive adenomas with the CAD-assisted colonoscopy. Although diminutive adenomas pose a lower risk of malignancy than the larger ones, the increase in ADR may eventually contribute to a decreased risk of interval colorectal cancer (CRC). However, due to the increase in PDR the CNN system significantly augmented the identification of hyperplastic polyps, which possibly links to additional unnecessary endoscopic resections and excessive workload.

Compared with CADe, under WLI endoscopy, advanced optical technologies including narrowband imaging (NBI) and endocytoscopy have been available for CADx in real-time performance.2, 4, 6 Mori et al.4 conducted a prospective study for a diagnose-and-leave strategy to identify diminutive, non-neoplastic recto-sigmoid polyps during on-going endocytoscopy, which provides 500-fold magnification. However, similar to the prior CNN application studies in the GI endoscopic field, the CAD-assisted endocytoscopy employed a creation of software referred to as EndoBrain (Olympus, Tokyo, Japan).2, 4 Another challenge of research is the application of the DL algorithm to the detection of proximal colonic lesions. In this regard, sessile serrated lesions should be the target of interest, because they are known to be associated with an alternative CRC pathway. Otherwise, WavSTAT4 (Spectra Science, San Diego, CA, USA) was approved by the US Food and Drug Administration as a CAD optical biopsy system.1

Both the AI systems under regulatory approval for optical biopsy have potential limitation of external validity in the earlier stages; selection bias has still not been eliminated. Extrapolating these findings obtained in the promising academic settings to community-based hospitals cannot be fully implemented in real practice. So far, operator dependency has been the difficult hurdle to overcome as such basic insertion and withdrawal skills are still required for colonoscopy. If AI is designed for NBI, endocytoscopy, or confocal laser endomicroscopy, it is still necessary to capture stable images for each advanced imaging module. The adaptabilities of the AI model on equipment manufactured by other vendors should also be explored. Of note, Yamada et al. have developed a vendor-free CNN model using the retrospectively accumulated still and video colonoscopy images which were captured from 30 endoscopists. Consecutively, their AI program has been learning about not only prominent but subtle lesions including slightly elevated and depressed lesions through the right and left colon and rectum, and are slated to start clinical trials using the AI model.6 Larger prospective multicenter trials are required to be conducted in diverse settings of hospitals, including endoscopists varying from non-experts to expert contributors, to provide such promising data in representative high volume centers.3, 4, 6

Currently, a facet of a recognized caveat of CNN is its black box nature where classifications of the DL algorithm are not readily apparent. Nowadays, CAD-assisted endoscopy is expected to serve as a second observer, helping to detect neoplastic lesions as early as possible. It is still impossible to employ purely AI-associated endoscopic devices in a true sense in real-world practice unless they gain the confidence of both patients and clinicians. Less considering state-of-art modular solutions such as optical biopsy, one may be moving towards the practical applicability of the AI models using routine WLI endoscopy for CADx as well as CADe, because the integration of AI technology to WLI endoscopic system would be convenient, speedy, objective, and user-friendly in daily use without its comprehensive advents. Tracking suspicious lesions needs to be performed with the real interest of screening to navigate automated diagnosis during on-going endoscopic performance using the familiar conventional observation of WLI. Although AI solutions are at a more advanced stage of development for colonic diseases than for conditions of the rest of the GI tract, there have been promising works of the AI application to upper GI endoscopy with WLI.7 The grid model on the DL-CNN basis to alert endoscopists to possible early gastric cancer and blind spots even on unprocessed videos would be extended for colonoscopy field.

This progress in notifying non-expert endoscopists proactively tracking suspicious lesions of neoplasia or indicating whether blind spots exist in real-time conventional endoscopy can be translated to the endoscopic training process and education, as observing the whole lumen is a basic prerequisite for qualified endoscopic diagnosis. Indeed, there are systemic standard protocols made to map the entire stomach during upper GI endoscopy, but these protocols are often not well followed and endoscopists may ignore some parts because of subjective factors or limited operative levels, which may lead to possible misdiagnosis. Incorporating the AI concept into a humanoid-robotic simulator for upper GI endoscopy8 can offer a possible self-training opportunity, as endoscopic learning is time- and cost-consuming.

Unrecognized and non-visualized lesions during endoscopy irrespective of CAD-assistance may be addressed with a combination of different technologies in the future. When assessing visibility of the colorectal lesions using the eye tracking method, the rate of missed diagnosis could be significantly lower, along with a shorter detection time.9 The AI-assisted tracking technology would play key roles to estimate not only the usefulness of endoscopic performance but also their educational effectiveness in training. On the other hand, the efforts for hypoxia imaging become increasingly fruitful when utilizing the AI system, which promises an exciting future of functional endoscopy. Furthermore, endoscopists should bear in mind that they should harmonize between the AI harness power and hyperspectral imaging that the human eye cannot appreciate or actually cannot see beyond the visible light region (nearly infrared) using methods such as photodynamic diagnosis.10

Since the CAD-assisted endoscopy system already possesses the capability of real-time decision making for treating colorectal polyps,2, 4, 5 the AI interpretation, hopefully along with uniform automated reporting, will enable doctors to help guide management to address the choice of precision treatment of GI cancers. Treatment options are slated to consist of diverse endoscopic resections, surgical interventions or systemic chemotherapy, molecular target therapy, or immunotherapy by incorporating CAD predilection of the tumor invasion depth, staging, and aggressive/indolent prognostic outcomes along with data sets based on clinical genetic sequencing in the future.

Self-driving automobiles are being used in real society. DL technology mimics the human brain; CNN resembles the human brain circuit; and similarly, automation of medical equipment such as automatic endoscopes would be created by imitating human procedures. In fact, self-propelling colonoscopies, soft-worm robotic endoscope, and capsule endoscope that allow for active locomotion in the small intestine and stomach using external magnetic force have been developed. The eye tracking system can also be useful to gain knowledge of experts’ behavior ultimately to optimize even automated endoscopic performance. In coming years, the next generation endoscopy would integrate robotic-assisted devices, eye tracking system, and AI technology.

Conflicts of Interest

Hajime Isomoto is a Deputy Editor-in-Chief of Digestive Endoscopy. The other authors declare no conflict of interests for this article.