Artificial intelligence and colonoscopy: the time is ripe to begin clinical trials

Recently, increased attention has been paid to optical biopsy modalities for diagnosing colorectal polyps. To reduce the cost of unnecessary polypectomies and/or pathological assessments of diminutive colorectal polyps (≤ 5 mm), adopting optical diagnoses based on an advanced endoscopic modality with high confidence prediction has been proposed as an alternative to histopathology by both the European Society of Gastrointestinal Endoscopy (ESGE) [1] and the American Society for Gastrointestinal Endoscopy [2]. However, to adhere to this initiative, endoscopists must exceed the performance thresholds for optical biopsy: > 90 % negative predictive value for identifying diminutive rectosigmoid adenomas is required to conduct “diagnose-and-leave” hyperplastic polyps, which is known as the Preservation and Incorporation of Valuable Endoscopic Innovations initiative-2 [PIVI-2]) [2]. However, these thresholds can only be met by experts and intensively trained endoscopists [3]. Standardized accreditation together with the auditing of endoscopists’ performance are also needed to put optical biopsy to practical use [4], and this has hindered its widespread use in routine colonoscopies.

Computer-aided diagnosis (CAD) powered by artificial intelligence is expected to provide a solution to these drawbacks of optical biopsy, because it potentially provides a standardized diagnostic performance for optical biopsy regardless of the endoscopists’ skill [5] [6]. The advantages of CAD are not limited to its performance. If regulatory bodies approve (accredit) its use, strict auditing of all endoscopists may not be needed or could be lessened, which would be a paradigm shift for optical biopsy. However, most of the previous studies on CAD for colonoscopy have focussed on advanced endoscopic imaging, such as magnified endoscopy, dual-focus endoscopy, and endocytoscopy, which are not widely available [5].

In this issue of Endoscopy, Fernández-Esparrach et al. demonstrate a significant advance in this clinical situation, as their CAD system used normal white light endoscopy as a target, which is undoubtedly the most widely available endoscopic modality, but one that has not been thoroughly examined in this field [7].

“As well as validating the efficacy of CAD, its practical role in routine colonoscopy should be discussed further and official guidance should be proposed to ensure patient safety: for example, should CAD serve as a second observer, a concurrent observer, or an independent decision maker?”

Their CAD model was developed based on the assumption that surface textural patterns guide endoscopists’ identification of dysplastic polyps. The developed CAD model automatically analyzed three metrics (contrast, tubularity, and branching) of the surface patterns of polyps to predict histology as either neoplastic or non-neoplastic based on machine learning. The performance of the model was retrospectively validated with a cross-validation method comprising 225 different polyp images. Histopathology of the resected specimens was used as the gold standard. The primary results were that the CAD system correctly classified 205 polyps as either neoplastic or non-neoplastic, with 92.3 % sensitivity and 89.2 % specificity based solely on white light endoscopic images. These results were comparable to those of expert endoscopists. Subanalysis revealed that diminutive polyps were correctly identified with an accuracy of 87.0 %, and that its negative predictive value for identifying diminutive rectosigmoid adenomas was 96.7 %, which met the PIVI-2 threshold. Although this subgroup analysis was of a relatively small number of polyps (n = 54), this achievement is notable because it is the first time the performance of a CAD system based on white light endoscopy exceeded the clinical threshold required for optical biopsy.

However, some limitations of this study hinder the generalization of the results. First, regions of interest (i. e. the location of polyps), which CAD should analyze, were manually indicated; therefore, fully automated diagnosis is not realized with the proposed method. This issue is common to most previously reported CAD systems [5], and is also pointed out as the primary drawback of current CAD systems by the ESGE technology review [1]. The additional limitation is that images with quality that was too poor to be analyzed by CAD were excluded from the study, and yet such cases comprise a substantial proportion of routine colonoscopies [8].

Considering that we have now obtained sufficient CAD options for optical biopsy that use not only advanced imaging but also normal white light endoscopy, it seems appropriate to open the door to the next stage: implementation into clinical practice. In fact, several groups have already submitted official applications for regulatory approval to use CAD for colonoscopy. However, before throwing artificial intelligence technology directly into the ocean of clinical practice, we should validate its performance through high quality prospective studies; to date, most of the relevant technologies have only been evaluated in experimental, nonclinical settings, which often leads to better results than in clinical practice. Prospective trials investigating real-time use of CAD during colonoscopy would provide information not only on its effectiveness and safety but also on its limitations, such as the presence of nonanalyzable images and/or prolonged procedure times [8]. Requirements for initiation of such prospective studies might include fully automated functionality (requiring no manual operation for segmentation of the target lesion) and high volume and highly varied learning images for machine learning to secure optimal performance of the system. As well as validating the efficacy of CAD, its practical role in routine colonoscopy should be discussed further and official guidance should be proposed to ensure patient safety: for example, should CAD serve as a second observer, a concurrent observer, or an independent decision maker? [9]. Furthermore, once its performance is widely evaluated in a prospective fashion with acceptable results, and its use is appropriately guided and controlled by regulatory bodies/authorities, CAD could become a “game changer” for optical biopsy during colonoscopy.

• References

• 1 East JE, Vleugels JL, Roelandt P. et al. Advanced endoscopic imaging: European Society of Gastrointestinal Endoscopy (ESGE) Technology Review. Endoscopy 2016; 48: 1029-1045
• 2 Rex DK, Kahi C, O’Brien M. et al. The American Society for Gastrointestinal Endoscopy PIVI (Preservation and Incorporation of Valuable Endoscopic Innovations) on real-time endoscopic assessment of the histology of diminutive colorectal polyps. Gastrointest Endosc 2011; 73: 419-422
• 3 Abu Dayyeh BK, Thosani N. ASGE Technology Committee. et al. ASGE Technology Committee systematic review and meta-analysis assessing the ASGE PIVI thresholds for adopting real-time endoscopic assessment of the histology of diminutive colorectal polyps. Gastrointest Endosc 2015; 81: 502.e1-502.e16
• 4 [Anonymous] National Institute for Health Care Excellence: Virtual chromoendoscopy to assess colorectal polyps during colonoscopy. 2017 https://wwwniceorguk/guidance/dg28
• 5 Mori Y, Kudo SE, Berzin TM. et al. Computer-aided diagnosis for colonoscopy. Endoscopy 2017; 49: 813-819
• 6 Byrne MF, Shahidi N, Rex DK. Will computer-aided detection and diagnosis revolutionize colonoscopy?. Gastroenterology 2017; 153: 1460-1464.e1
• 7 Sánchez-Montes C, Sánchez FJ, Bernal J. et al. Computer-aided prediction of polyp histology on white light colonoscopy using surface pattern analysis. Endoscopy 2018; 51: 261-265
• 8 Mori Y, Kudo SE, Misawa M. et al. Real-time use of artificial intelligence in identification of diminutive polyps during colonoscopy: a prospective study. Ann Intern Med 2018; DOI: 10.7326/m18-0249.
• 9 East JE, Rees CJ. Making optical biopsy a clinical reality in colonoscopy. Lancet Gastroenterol Hepatol 2018; 3: 10-12