Human-machine collaboration: bringing artificial intelligence into colonoscopy

Colorectal cancer (CRC) is the third most common malignancy worldwide. Colonoscopy offers protection against the development of CRC by detection and resection of neoplastic lesions. Unfortunately, the procedure remains highly operator dependent. A meta-analysis of tandem colonoscopy studies revealed a pooled miss rate of 22% for polyps of any size.¹ Postcolonoscopy CRCs are associated with low adenoma detection rates (ADR) and incompletely resected or missed lesions are recognised as key contributory factors.² International efforts to improve quality must be commended, particularly those led by the Joint Advisory Group on Gastrointestinal Endoscopy in the UK, where individual colonoscopy performance is assessed by quality assurance using key performance indicators. Despite these efforts further significant improvement is needed.

Numerous strategies have been used in attempts to improve ADRs including educational interventions, enhanced imaging techniques and mechanical devices to improve mucosal exposure. Computer-aided detection and diagnosis (CAD) systems using advanced artificial intelligence (AI) techniques represent an emerging technology that will likely lead to a paradigm shift in the field.³

Machine learning is a type of AI applied to systems, which allows for automatic learning and improved performance on data sets without the need for explicit programming of prediction rules. Advanced deep-learning-based approaches have revolutionised the area by utilising artificial neural networks.⁴ These systems are biologically inspired by the concept of neurons and synapses in the human brain to discover image features that optimally represent the data for a specific task. Examples of successful applications include demonstrations that deep-learning algorithms are able to match the performance of expert dermatologists in differentiating benign from malignant skin lesions and are better than pathologists at assessing slide images for the presence of breast cancer metastases within axillary node specimens during a timed exercise designed to simulate clinical practice.^{5 6}

CAD systems are being developed rapidly for real-time polyp detection. A recent study by Urban et al describes the use of a deep convolutional neural network (CNN) trained on 8641 images from 2000 patients, leading to a polyp detection accuracy of 96.4% in an independent set of 1330 images containing 672 polyps operating at real-time video rate.⁷ More importantly, when further evaluated on nine retrospectively collected colonoscopy videos, an additional 17 polyps that were not removed at the index procedure were identified using CNN assistance with a relatively low false positive rate.

Furthermore, some CAD systems can now match human experts in performing ‘optical biopsies’.⁸ In 2017, the National Institute for Health and Care Excellence recommended that virtual chromoendoscopy can be used instead of histopathology to determine if diminutive (≤5 mm) polyps are adenomatous or hyperplastic under certain conditions which includes high confidence assessments.⁹ Studies have demonstrated that the Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI) standards set for such a ‘resect and discard’ strategy can be achieved in academic settings but not always in community-based practice, limiting its incorporation into routine care.¹⁰ CAD offers a promising timely solution to overcome this barrier by providing decision support. Byrne et al used a CNN to differentiate diminutive adenomas from hyperplastic polyps on unaltered narrow band imaging (NBI) videos collected from a previous study using standard colonoscopy.¹¹ The algorithm provided an associated probability score for predictions. Using histopathology as gold standard, the model provided real-time high confidence decisions for 106 diminutive polyp videos, the overall accuracy achieved was 94% (sensitivity 98%, specificity 83%, positive predictive value 90%, negative predictive value 97%). More recently, Mori et al conducted a prospective study using a CAD system for endocytoscopy, which produces ultrahigh magnification images that allow for in vivo cellular or microvascular assessment with staining or NBI, respectively.¹² The system achieved the PIVI threshold for a ‘diagnose and leave’ strategy for diminutive, non-neoplastic, rectosigmoid polyps. The authors commented on a future clinical trial to investigate the CAD system for a ‘resect and discard’ strategy.

Clinical trials utilising AI-powered software during endoscopy are imminent and it is vital that clinical endoscopists are engaged at this time of rapid technological development. It is possible that new quality metrics will emerge based on computational advances. An advantage of CAD is the potential for relatively simple integration with standard endoscopy equipment by using ‘plug-in’ software. However, before implementation into routine National Health Service practice or bowel cancer screening programmes can even be considered there must be further evaluation of clinician acceptance and assessment of potential impact on workflow. There are also regulatory hurdles that must be overcome including addressing concerns about the lack of transparency and complexity involved in deep-learning-based decision-making which is often considered a ‘black box’. Despite these challenges, we should take this opportunity to reflect on the quality of our own current endoscopy practice and embrace a future ‘human-machine collaboration’ that will pave the way for the very highest quality endoscopy for our patients.