Answers are grounded in this paper and cite their sources.
We stand at the threshold of a new era: the incorporation of artificial intelligence (AI) into routine gastrointestinal (GI) endoscopy practice to improve clinical performance. In this issue of Journal of Gastroenterology and Hepatology, review articles will address in depth the role of AI in medical image analysis in general and the use of AI in upper endoscopy, small bowel capsule endoscopy (CE), colonoscopy, and pancreatico-biliary endoscopy in particular.1-5 The clinical applications of AI relevant to GI endoscopy are computer-aided detection (CADe) and computer-aided diagnosis (CADx). Current AI algorithms are based on deep learning, where a hierarchy of linear transformations combined with nonlinear functions is structured together in complex layers to form convolutional neural networks (CNNs). Early layers of the network extract low-level features (edges and corners) and deeper layers extract high-level features (complex visual objects, such as faces) based on the low-level features from the first layers. The classification is usually based on a combination of the high-level features using fully connected layers. A larger number of layers form a deeper network, which has a higher representation capability than a shallow network. The focus of CADx systems is on obtaining accurate classification results and correct diagnoses. AI models should include well-calibrated classification and reliable uncertainty estimates to indicate the likelihood the classification or diagnosis is correct, and the reliability of the result. Such parameters will provide clinicians the basis to either accept or overturn the assessment based on AI.6 It must be recognized that there are innate differences in the design and algorithms of different AI systems. AI algorithms may be developed using specific endoscopy technologies and settings, such as white light imaging (WLI), narrowband imaging (NBI) with or without optical magnification, and endocytoscopy, and thus will be valid only when applied in the same setting. Training and testing of the AI algorithm in more than one population will enhance its robustness and facilitate generalization to different clinical settings. There is a branch in machine learning called lifelong learning, which is about the development of systems that can continuously learn, but it is something more for the future.7 The quality of the endoscopy procedure and the findings of endoscopy are both important parameters that must be addressed. Using AI to assess and document the quality of the procedure serves two purposes: it can facilitate clinical audit of the clinical performance of endoscopists retrospectively and it can also enhance the quality of the procedure by modifying behavior and technique real time. Early data suggest that real-time AI assessment of procedural quality is feasible, such as assessing the thoroughness of upper GI endoscopy examination8 and assessing the quality of colonoscopy in terms of bowel preparation adequacy, and timing and stability of colonoscopy withdrawal.9 Most studies of AI in endoscopy have focused on lesion detection and characterization. The most robust data are in the context of colonoscopy, in particular for polyp detection, where randomized controlled trials have been published.4 The data for upper GI endoscopy are less robust than colonoscopy, with most studies based on retrospective reviews of selected still images.2 In the context of small bowel CE, there are already AI algorithms integrated into the system to improve the efficiency of capsule reading, such as the Suspected Blood Indicator and Quick View algorithms (Medtronic). These are based on classic machine learning rather than CNN and are still not able to significantly shorten CE reading time due to limitations in diagnostic accuracy. Preliminary published data of CE AI algorithms developed using CNN are very promising.3 There is a paucity of published data concerning the use of AI in pancreatico-biliary endoscopy. This presents significant opportunities for further research.5 Recently, Marya et al. reported the use of endoscopic ultrasound-based CNN to differentiate autoimmune pancreatitis (AIP) from pancreatic adenocarcinoma (PAC), chronic pancreatitis (CP), and normal pancreas (NP). The reported sensitivity and specificity to distinguish AIP from NP were 99% and 98%, to distinguish AIP from CP were 94% and 71%, and to distinguish AIP from PAC were 90% and 93%, respectively.10 The hope is for AI to overcome human limitations and enhance our performance in diagnostic endoscopy. Specifically, to overcome inconsistencies by endoscopists in detection rates and diagnostic accuracies due to inexperience or fatigue with loss of focus. AI has the potential to facilitate acquisition of expertise by providing positive feedback on the quality of endoscopy real time and highlighting pathology and the likely diagnosis. The impact on the expert endoscopist is less clear, and indeed, concerns about deskilling due to overreliance on AI have been expressed. But the hope is that even the competency of the expert endoscopist can be enhanced. There are challenges that must be addressed if we are to fully embrace AI and maximize its potential. In terms of clinical utility, data about test performance such as sensitivity, specificity, and accuracy alone are inadequate. It is also crucial to establish the clinical impact to patient care. It may not be very relevant if AI-assisted colonoscopy merely detects more diminutive polyps which may not progress. Of greater impact would be the detection of sessile serrated lesions and lateral spreading tumors that may be easily missed and that are of greater clinical impact due to the higher risk of disease progression. This is also relevant for upper endoscopy, where the possibility of missing early GI cancer despite meticulous examination remains a concern. It must be appreciated that there may be a ceiling in terms of clinical impact which is dependent on the source technology used to create the AI algorithm. AI algorithms using endocytoscopy which has much higher degree of magnification than magnifying NBI would be expected to outperform an AI algorithm based on a lower degree of magnification for diagnostic evaluation.11 To break the barrier in terms of further improving diagnostic accuracy, alternative solutions may be needed. Regulatory concerns and medicolegal liabilities, especially if the AI makes a wrong assessment and the physician acts on it, or vice versa, are other issues that must be addressed. Similar to other fields of medicine, such as in breast cancer screening,12 there is a need for independent assessment of commercial AI systems. External evaluations tend to show a more realistic picture about the actual performance of such systems. There is tremendous potential for AI to enhance endoscopist performance and improve efficiency. The tentative first steps towards mainstream clinical practice have started, but there are hurdles to overcome for its full potential to be harnessed. Nil.
No takes yet. Share an insight, caveat, or question.
Ang et al. (2021) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: