Artificial Intelligence in Colonoscopy: Where Are We Now in 2024?

Abstract

Introduction

Colonoscopy has a crucial role in reducing colorectal cancer incidence and mortality. Different artificial intelligence (AI) systems were developed to further improve its quality assurance (computer-aided quality improvement [CAQ]), lesion detection (computer-aided detection [CADe]), and lesion characterization (computer-aided characterization [CADx]). There were studies investigating the roles of these AI systems in different domains of standard colonoscopies.

Methods

In this state-of-the-art narrative review, we summarize the current evidence, discuss existing limitations, as well as explore the future directions of AI in colonoscopy.

Results

CAQ enhances colonoscopy quality through real-time feedback and quality monitoring systems, but the studies have inconsistent results due to small training datasets and varied methodologies. CADe increases adenoma detection rate and reduces adenoma missed rates, but there are concerns about false positives, unnecessary polypectomies, potential deskilling of endoscopists, and cost-effectiveness. CADx systems have mixed results and accuracies in differentiating polyp types, and its use is further hindered by inadequate representation of sessile serrated lesions and a lack of rigorous trials comparing it with standard colonoscopy.

Conclusion

Despite the emerging evidence of AI-assisted colonoscopy, its potential drawbacks and limitations may hinder the further implementation in real-world clinical practice. Long-term data on clinical efficacy, cost-effectiveness, liability, and data sharing are the key areas to be addressed.

Introduction

Colorectal cancer (CRC) is the third most prevalent cancer and the second leading cause of cancer-related death in the world [1, 2]. Colonoscopy reduces the CRC incidence and mortality by removing pre-malignant polyps and early intervention of CRC [3–6]. Traditionally, adenoma detection rate (ADR), which is defined by the proportion of average-risk patients with at least one adenoma detected among screening colonoscopies, has been shown to be inversely associated with interval CRC [7, 8]. However, a considerable proportion of adenomas may be missed during colonoscopy [9]. Some of the missed adenomas were overlooked due to suboptimal quality during colonoscopy, inadequate mucosal exposure, and cognitive limitations on the part of the endoscopists [9–11]. Moreover, inaccurate lesion diagnosis and characterization may lead to incorrect endoscopic treatment and inadequate resection. With the rapid development of artificial intelligence (AI), different computer-aided tasks are gaining attention as reliable tools to overcome these challenges in colonoscopy. However, these new technologies also have limitations and areas needing improvement before full implementation. In this article, we review the current evidence, discuss potential drawbacks and limitations, as well as explore the future directions of AI in colonoscopy.

Methods

This is a narrative review of the literature on the status of AI applications for quality control, polyp detection, and characterization in colonoscopy. The narrative review design is selected as it offers a broader scope of synthesis than the systematic review design, which would enable the fields mentioned and future directions for AI use in colonoscopy to be summarized and discussed.

A comprehensive literature search was conducted on PubMed, Embase, and Medline databases from inception to and including March 31, 2024. To obtain the most relevant sources, the combination of search terms was used with Boolean operators “AND” and “OR”: “colonoscopy,” “artificial intelligence,” “quality control,” “computer-aided detection,” “computer-aided diagnosis.” Electronic searches were supplemented with manual searches of references of retrieved studies to identify other relevant publications. Only studies published in English were included in this review article. The searching workflow is shown in Figure 1.

Fig. 1.

Literature search process.

Results

Current Evidence

Strengths and Opportunities

Computer-Aided Quality Improvement (CAQ) Control. CAQ (Fig. 2) was developed to improve the quality of colonoscopy examinations. Apart from ADR, other validated quality indicators of colonoscopy include bowel preparation adequacy rate, caecal intubation rate, and withdrawal time. At present, there are a number of AI systems that have been developed with the aim to target these quality indicators.

Fig. 2.

CAQ – ENDOANGEL (Wuhan ENDOANGEL Medical Technology Co., Ltd) records the insertion and withdrawal time, estimated BBPS, speed of withdrawal, and location of the colonoscope.

ADR is the most clinically relevant quality indicator, yet not the most perfect as it does not take into account the potential missing polyp(s) behind the colonic folds. Polyps may be hidden in blind spots or behind the colonic folds, and they may be missed without deliberate examination of these areas. Liu et al. [12] proposed an AI system focusing on providing real-time feedback on completeness of colon examination. This significantly increases fold examination quality (FEQ) in endoscopists with low ADR (0.29 vs. 0.23, p < 0.001), as compared to the control group without AI assistance. They used the FEQ scoring criteria developed by Duloy et al. [13] to further assess the FEQ: score 0 = very poor performance (not looking behind any folds, “straight pull-back” technique), 1 = poor, 2 = fair, 3 = good, 4 = very good, and 5 = excellent performances (looking behind all folds to allow for ideal mucosal visualization). It is believed that this system helps improve ADR by providing real-time feedback on completeness of colon examination. Chang et al. [14] have also created a deep learning (DL) algorithm which is a high-quality validated photo documentation of polyp detection, caecal intubation, and bowel preparation. This DL algorithm aimed to provide automated assessment of and timely feedback on the quality of colonoscopy which otherwise is almost not attainable with manual review. The accuracy of DL is 95.45%, 98.46%, and 97.56% in identifying caecal, instrument, and retroflexion images, respectively, and 94.57% in differentiating bowel preparation status.

On the other hand, the ENDOANGEL was designed to provide automated monitoring of bowel cleanliness. The feedback relayed to the endoscopist has prompted more frequent washing of colon and therefore more likely to increase the exposure of mucosa. A randomized controlled trial (RCT) conducted by Gong et al. [15] showed that the ADR in the ENDOANGEL group was significantly higher (16% vs. 8%), as compared to the control (unassisted colonoscopy) group. This was also supported by a prospective observational study by Zhou et al. [16] who noted that there was a significant inverse correlation between the automated Boston Bowel Preparation Score (e-BBPS) and ADR (Spearman’s rank −0.976, p < 0.010).

Furthermore, adequacy of inspection time cannot be overemphasized, and latest American Society for Gastrointestinal Endoscopy (ASGE) 2024 guideline [17] has recommended withdrawal time of at least 8 min. Su et al. [18] have conducted an RCT on the automatic quality control system which supervises scope withdrawal stability and polyp detection. The study showed that automatic quality control system has improved mean withdrawal time from 5.68 min to 7.03 min (p < 0.001) by generating audio prompts for endoscopists to slow down when unstable frames were displayed, or suboptimal bowel preparation with Boston Bowel Preparation Scale (BBPS) <2 was detected.

CAQ system was also found by Yao et al. [19] to significantly augment the performance of computer-aided detection (CADe) when they are used in combination (COMBO). The ADR of COMBO and CADe was 30.6% and 21.27% (p = 0.024) respectively. This is likely attributed to more thorough examination of mucosa by CAQ.

Computer-Aided Detection. Many RCTs and meta-analyses have demonstrated significant benefits of CADe-assisted colonoscopy (Fig. 3) compared to standard colonoscopy (SC) in increasing ADR, regardless of colonoscopy indication and endoscopist experience. There have been over 20 RCTs since 2019 investigating the role of CADe on ADR (Table 1), with majority showing significant superiority over SC. In a meta-analysis, Lou et al. [20] found that AI-aided systems experienced a 24.2% relative increase (7.4% absolute risk difference) in ADR. In another meta-analysis, Hassan et al. [21] also showed a significantly higher ADR in CADe group than SC (44.0% vs. 35.9%). This enhanced performance was consistently observed in both experienced and junior endoscopists in different trials.

Fig. 3.

CADe – ENDO-AID (OIP-1; Olympus, Tokyo, Japan) detects polyp with on-screen alert.

Table 1.

Parallel RCTs of CADe vs. SC

First author [ref.] (year)	Sample size (CADe vs. SC)	CADe programme	ADR, % (AI vs. SC)	Absolute ADR gain, %	Statistically significant superiority (Y/N)
Wang et al. [22] (2019)	522 vs. 536	EndoScreener	29.1 vs. 20.3	8.8	Y
Gong et al. [15] (2020)	355 vs. 349	ENDOANGEL	16 vs. 8	6	Y
Liu et al. [23] (2020)	393 vs. 397	EndoScreener	29.01 vs. 20.91	8.1	Y
Liu et al. [24] (2020)	508 vs. 518	Local system	39.1 vs. 23.9	15.2	Y
Repici et al. [25] (2020)	341 vs. 344	GI Genius	54.8 vs. 40.4	14.4	Y
Su et al. [18] (2020)	308 vs. 315	Local system	28.9 vs. 16.5	12.4	Y
Wang et al. [26] (2020)	484 vs. 478	EndoScreener	34 vs. 28	6	Y
Shen et al. [27] (2021)	64 vs. 64	Local system	53.1 vs. 29.7	23.4	Y
Repici et al. [28] (2022)	330 vs. 330	GI Genius	53.3 vs. 44.5	8.8	Y
Rondonotti et al. [29] (2022)	405 vs. 395	CAD EYE	53.6 vs. 45.3	8.3	Y
Shaukat et al. [30] (2022)	682 vs. 677	Local system	47.8 vs. 43.9	3.9	N
Yao et al. [19] (2022)	268 vs. 271	ENDOANGEL	21.27 vs. 14.76	6.51	Y
Ahmad et al. [31] (2023)	308 vs. 306	GI Genius	71.4 vs. 65.0	6.4	N
Aniwan et al. [32] (2023)	312 vs. 310	CAD EYE	52.2 vs. 41.9	10.3	Y
Engelke et al. [33] (2023)	122 vs. 110	GI Genius	33.6 vs. 18.1	15.5	Y
Gimeno-Garcia et al. [34] (2023)	155 vs. 157	ENDO-AID	55.1 vs. 43.8	11.3	Y
Hüneburg et al. [35] (2023)	50 vs. 46	CAD EYE	36 vs. 26.1	9.9	N
Karsenti et al. [36] (2023)	1,003 vs. 1,012	GI Genius	37.5 vs. 33.7	3.8	N
Lachter et al. [37] (2023)	330 vs. 344	DEEP2	37 vs. 27	10	Y
Mangas-Sanjuan et al. [38] (2023)	1,610 vs. 1,603	GI Genius	64.2 vs. 62	2.2	N
Vilkoite et al. [39] (2023)	194 vs. 206	ENDO-AID	30.41 vs. 20.87	9.54	N
Wang et al. [40] (2023)	635 vs. 625	EndoScreener	25.8 vs. 24.0	1.8	N
Wei et al. [41] (2023)	387 vs. 382	EndoVigilant	35.9 vs. 37.2	−1.3	N
Wei et al. [42] (2023)	124 vs. 120	EndoVigilant	68.5 vs. 80	−11.5	N
Xu et al. [43] (2023)	1,519 vs. 1,540	Eagle-Eye	39.9 vs. 32.4	7.5	Y
Desai et al. [44] (2024)	509 vs. 522	CAD EYE	46.8 vs. 42.9	3.9	N
Lui et al. [45] (2024)	238 vs. 214	Olympus OIP-1	53.8 vs. 46.3	7.5	Y
Lau et al. [46] (2024)	386 vs. 380	ENDO-AID	57.5 vs. 44.5	13	Y
Miyaguchi et al. [47] (2024)	400 vs. 400	CAD EYE	58.8 vs. 43.5	15.3	Y
Schöler et al. [48] (2024)	122 vs. 118	CAD EYE/Medtronic GI Genius	43 vs. 41	2	N
Tiankanon et al. [49] (2024)	400 vs. 400	DeepGI	54.8 vs. 38.3	16.5	Y
	400 vs. 400	CAD EYE	50 vs. 38.3	11.7	Y

CADe, computer-aided detection; SC, standard colonoscopy; ADR, adenoma detection rate.

Studies also showed that CADe could reduce adenoma missed rate (AMR). Most tandem RCTs involving two endoscopic passes reported a significant AMR reduction with the use of CADe over SCs (Table 2). The meta-analysis from Lou et al. [20] revealed that individuals who intervened with CADe experienced a 50.5% relative decrease (17.5% absolute risk difference) in AMR. Hassan et al. [21] also found a lower AMR when the first colonoscopy was done with CADe (16% vs. 35% with SC), with a risk ratio of 0.45 and moderate level of heterogeneity (I² = 49%). Moreover, Yao et al. [50] revealed that AI-assisted colonoscopy lowered the AMR of novices, making them non-inferior to experts.

Table 2.

Tandem RCTs of CADe vs. SC

First author [ref.] (year)	Sample size (CADe vs. SC)	CADe programme	AMR % (CADe vs. SC)	Statistically significant superiority (Y/N)
Wang et al. [51] (2020)	184 vs. 185	EndoScreener	13.89 vs. 40	Y
Kamba et al. [52] (2021)	178 vs. 177	Local system	13.8 vs. 36.7	Y
Glissen et al. [53] (2022)	113 vs. 110	EndoScreener	20.12 vs. 31.25	Y
Wallace et al. [54] (2022)	116 vs. 114	GI Genius	15.5 vs. 32.4	Y
Lui et al. [55] (2023)	108 vs. 108	Local system	20 vs. 14	N
Nakashima et al. [56] (2023)	207 vs. 208	CAD EYE	11.9 vs. 26	Y
Yamaguchi et al. [57] (2023)	113 vs. 118	CAD EYE	25.6 vs. 38.6	Y
Yao et al. [50] (2023)	227 vs. 458	ENDOANGEL	18.82 vs. 43.69	Y
Maas et al. [58] (2024)	61 vs. 66	MAGENTIQ-COLO	19 vs. 36	Y

CADe, computer-aided detection; SC, standard colonoscopy; AMR, adenoma missed rate.

Apart from ADR and AMR, studies also revealed the effect of CADe in other important parameters, including proximal adenoma detection, adenoma per colonoscopy (APC), and polyp detection rate (PDR). Lesions in the proximal colon (i.e., from caecum to transverse colon) are more frequently missed during colonoscopy [59]. Lui et al. [55] showed that CADe enhanced proximal adenoma detection in patients with fair bowel preparation, shorter withdrawal time, and endoscopists with lower ADR. Engelke et al. [33] found that ADR increase was particularly strong for the detection in elderly patients aged ≥50 years. Lou et al. [20] observed a significant increase in PDR, PPC, and APC, along with a significant decrease in polyp miss rate, and they also found that endoscopists with lower ADR (or PDR), shorter inspection time, patients with lower BMI, and younger age will benefit more from AI-aided colonoscopy.

Computer-Aided Characterization. The American Society for Gastrointestinal Endoscopy (ASGE) Technology Committee has set the benchmarks (i.e., preservation and incorporation of valuable endoscopic innovation thresholds) for which computer-aided characterization (CADx) systems (Fig. 4) would have to achieve before setting the stage for their applicability in the clinical setting. CADx needs to demonstrate a negative predictive value (NPV) of at least 90% for adenomas, especially when assessing recto-sigmoid colon polyps less than or equal to 5 mm. This has led to several studies aiming to validate the existing CADx systems against these benchmarks (Table 3) and a potential “diagnose-and-leave” strategy.

Fig. 4.

CADx – ENDOBRAIN (Cybernet Systems Corp., Tokyo, Japan) characterizes lesions into non-neoplastic and neoplastic.

Table 3.

Studies on CADx

First author [ref.] (year)	CADx programme	Polyps, n	NPV of combined CADx + all endoscopists	NPV of CADx with experts	NPV of CADx with non-experts
Mori et al. [60] (2018)	EB-01 prototype (Cybernet Systems)	466	96.4% (95% CI: 91.9%–98.8%)	91.8% (95% CI: 88.0%–94.7%)a	86.6% (95% CI: 82.1%–90.3%)b
Rondonotti et al. [61] (2022)	CAD EYE (Fujifilm Co., Tokyo, Japan)	596	91.0% (95% CI: 87.1%–93.9%)	92.4% (95% CI: 87.5%–95.5%)	88.6% (95% CI: 80.5%–93.6%)
Hassan et al. [62] (2022)	GI Genius Intelligent Endoscopy Module (version 3.0.0; Medtronic)	291	97.6% (95% CI: 94.8%–99.1%)	97.6% (95% CI: 94.8%–99.1%)	(Not available)
Barua et al. [63] (2022)	EndoBRAIN (Cybernet Systems Corp., Tokyo, Japan)	892	92.8% (95% CI: 90.1%–94.9%)	(Not available)	92.8% (95% CI: 90.1%–94.9%)
Minegishi et al. [64] (2022)	NBI-CAD	465	96.4% (95% CI: 91.9%–98.8%)	(Not available)	(Not available)
Li et al. [65] (2023)	CAD EYE (Fujifilm Co., Tokyo, Japan)	661	58.6% (95% CI: 53.5%–63.6%)c	63.8% (95% CI: 54.4%–72.5%)d	55.0% (95% CI: 47.2%–62.6%)e
Houwen et al. [66] (2023)	POLAR system	423	46.3% (95% CI: 34.3%–58.2%)f	(Not available)	(Not available)
Hassan et al. [67] (2023)	CAD EYE (Fujifilm Co., Tokyo, Japan) and GI Genius (version 3.0.0; Medtronic)	325	97.7% (95% CI: 96.0%–99.5%)	(Not available)	(Not available)

CADx, computer-aided characterization; NPV, negative predictive value.

^aNPV of expert endoscopists only (without CADx).

^bNPV of non-expert endoscopists only (without CADx).

^cNPV of CADx only (performed by both expert and non-expert endoscopists).

^dNPV of CADx only (performed by both expert endoscopists).

^eNPV of CADx only (performed by both non-expert endoscopists).

^fNPV of CADx only.

Currently, studies on CADx have yielded conflicting results. In 2022, Hassan et al. [62] employed the GI Genius CADx system to differentiate 291 recto-sigmoid colon polyps into adenomatous versus non-adenomatous polyps. The GI Genius system alone was able to achieve a NPV of 97.6%. This was similar to the NPV of 97.6% achieved when the assessment was performed by GI Genius and endoscopists [62]. In the same year, Barua et al. [63] tested the EndoBRAIN system to differentiate 892 recto-sigmoid colon polyps as neoplastic versus non-neoplastic in nature. The EndoBRAIN system achieved an NPV of 92.8% as compared to 91.5% by endoscopists alone [63]. Separately, Hassan et al. [67] compared 2 CADx systems (i.e., CAD EYE and GI Genius) to differentiate 325 recto-sigmoid colon polyps into adenomatous versus non-adenomatous polyps. They were able to demonstrate an NPV of 97.0% for CAD EYE and 97.7% for GI Genius [67].

On the other hand, Rondonotti et al. [61] in 2022 challenged the CAD EYE system in real time to differentiate adenomatous from non-adenomatous polyps resected from the recto-sigmoid colon. CAD EYE alone achieved an NPV of 86.7%, which was slightly lower as compared to the NPV of 90.9% achieved by the endoscopists alone. In a similar fashion, CAD EYE was validated in real time again by Li et al. [65] in the following year for all polyps in the colon. This time, CAD EYE alone demonstrated an NPV of 58.6%, as compared to 63.4% for the endoscopists. Separately, Houwen et al. [66] designed a new CADx system (POLAR) in 2023 and performed a validation study on 423 colonic polyps to differentiate them into neoplastic versus non-neoplastic diminutive lesions. The NPV for their new system was 46.3%, compared to 58.1% for endoscopists alone.

Interestingly, the clinical impact of integrating narrowband imaging (NBI) or endocytology into CADx was also assessed by Minegishi et al. [64] and Mori et al. [60] respectively. In the former study, the NBI-CAD system was used to differentiate colonic polyps into either hyperplastic, sessile serrated lesion (SSL), or neoplastic/adenomatous polyp. No difference in the sensitivity between NBI-CAD (93.3%) and endoscopists (94.4%) was observed [64]. In the latter study, Mori and colleagues were able to integrate endocytoscope into CADx to predict neoplastic versus non-neoplastic colon polyps. This technique was able to achieve an NPV of 93.7%–96.4% for recto-sigmoid polyps and 60% for polyps located in the rest of the colon [60].

Drawbacks and Limitations

Quality Control. Although AI shows promising results in quality control, there are several limitations in hindering its use in clinical practice. The evidence to date supporting CAQ during colonoscopy is conflicting, in part due to the lack of adequately powered studies, small training datasets for training for some models, and a relatively smaller number of available CAQ systems compared to CADe and CADx [68].

For instance, during a high procedural volume to achieve an accurate estimation of ADR, the number of colonoscopies per endoscopist in some CAQ studies evaluating withdrawal speeds and time may be too low to accurately detect a change in ADR [50, 69, 70]. This may explain the differing results seen in CAQ studies measuring quality of colonoscopy withdrawals against ADR [18, 19, 71]. Furthermore, the risk of overfitting when AI systems are trained on small datasets also limits the applicability of such systems in clinical practice [72]. In the study by Liu et al. [12], the deep convoluted neural network was initially trained on images derived from only 25 colonoscopy videos and was validated on 103 colonoscopy videos from 11 endoscopists in a span of 3 months. This calls into question the generalizability of the FEQ assessments made by the CAQ system, since the system performance relies heavily on the volume and quality of the datasets used in the training phase. The indications and drawbacks of various AI systems are summarized in Table 4.

Table 4.

Indications and drawbacks of various AI systems on colonoscopy quality control

AI algorithm	Indication	Result	Drawback
FEQ	Provides feedback on completeness of examination	AI 0.29 vs. control 0.23, p < 0.001	Prospective observational study with small sample Threshold of the system’s FEQ is not determined
DL algorithm	Reviews polyp detection, caecal intubation, and bowel preparation and compares against an electronic database of high quality and validated photo documentation of scope images	Accuracy rate	Uses images from a single system only, i.e., Olympus EVIS LUCERA ELITE Uses static images which may not be applicable in real-time situation with peristalsis
		Caecal intubation: 95.45%
		Rectal retroflexion: 97.56%
		Differentiating bowel preparation status: 94.57%
ENDOANGEL	Monitors bowel cleanliness	ADR of AI 16% vs. control 8%, p = 0.001	Contains more examples of adequate bowel preparations than inadequate bowel preparations, which may have introduced bias The algorithm learns that longer time spent in procedure equates to inadequate bowel preparation, which may yield inaccuracies
AQCS	Supervises scope withdrawal stability and time	Mean withdrawal time improves from 5.68 to 7.03 min, p < 0.001	RCT is limited to a single endoscopic centre Training only on single system, i.e., Pentax

FEQ, fold examination quality; DL, deep learning algorithm; AQCS, automatic quality control system.

Computer-Aided Detection. Despite the strong evidence generated from clinical trials, there are several potential drawbacks which need to be addressed before real-world utilization. Firstly, although CADe increases ADR, most studies did not demonstrate an increase in the detection rates of advanced or large (>10 mm) adenomas, SSLs, and CRC [20, 21, 38, 73]. The increased rate of detecting non-neoplastic polyps may lead to higher rate of unnecessary polypectomies. Most of these polypectomies were likely due to small hyperplastic polyps that were not associated with interval CRC development [74, 75].

Secondly, CADe may introduce extra-burden during endoscopy. Lou et al. [20] reported false-positives rates from CADe ranged between 7.1% and 20.1% from 8 RCTs. The false-positive signals may increase the psychological distraction of the endoscopists, leading to a longer inspection and withdrawal time [22, 76].

Thirdly, an overreliance on CADe may also result in deskilling of endoscopists. Troya et al. [77] evaluated the reaction time for polyp detection and eye-tracking metrics of endoscopist with and without CADe, which showed that use of CADe did not improve human reaction times. It increased misinterpretation of normal mucosa and decreased the eye travel distance. Possible consequences might be prolonged examination time and deskilling [77].

Finally, besides clinical efficacy, cost-effectiveness is another important area to be considered. As mentioned above, CADe increases PDR, which leads to the increasing number of polypectomies and subsequent histopathological examinations. A simulation study suggested with the use of CADe the number of intensive surveillance colonoscopies may increase by 35% in the USA [78]. A microsimulation study suggested that the use of CADe can increase health-care cost in short term, but it could contribute to the reduction of CRC incidence and death, which could save the overall health-care costs in the long term instead [79]. In this regard, the 2023 WEO position statement recommended health-care delivery systems and authorities to evaluate the cost-effectiveness of CADe to support its use in clinical practice [80].

Computer-Aided Characterization. Current available studies are limited by several factors. Firstly, all SSLs are either under-represented or misrepresented in their studies. SSLs, unlike adenomatous lesions, represent a different malignant potential. However, this was left out of the analysis by Mori et al. [60] and Li et al. [65], while Rondonotti et al. [61] classified them under non-adenomatous lesions [61, 65, 60].

Secondly, all the studies currently aim to validate CADx only as a diagnostic tool. However, CADx has yet to be tested against SC in an RCT setting to demonstrate a net benefit.

Thirdly, cost-effectiveness has been cited as one of the benefits that CADx can potentially provide. This is because of its complimentary role as described above for endoscopists to implement the “diagnosed and leave” strategy, removing the need for a further histopathological diagnosis of diminutive polyps. However, none of the studies have quantified the cost savings of CADx from that aspect. Cost-effectiveness would need to be justified before its implementation and integration into our routine endoscopic workflow.

Finally, current studies did not focus on using post-colonoscopy colorectal cancer (PCCRC) as a study outcome. PCCRC is recognized as the most ideal outcome measure for studies focusing on polyp surveillance. Future long-term studies should focus on measuring PCCRC directly as a primary outcome of CADx trials to confirm its role in the clinical setting.

Real-World Implementation

Quality Control

At present, CAQ still limited to the research setting and not ready for clinical use. AI systems require more validation and on a larger scale, especially across different institutions, patient profiles, endoscopy systems, and different skill levels of endoscopists. For example, the AI algorithm developed by Peterson et al. [81] demonstrated the inaccuracy in identifying the number of polyps although it has showed a high accuracy in recommending the surveillance interval based on colonoscopy and pathology report which confirms adenoma. This demonstrated that AI systems are not ready for the current clinical climate, yet much work is needed to improve the systems to produce consistent results.

Computer-Aided Detection

To date, CADe use is still largely limited in research settings, and its performance of implementation into clinical setting is still unclear. Only a few studies focused on this area. Nehme et al. [82] implemented CADe systems in endoscopy rooms and allowed endoscopists to decide whether to activate it, resulting in activation in only 52.1% of cases. In addition, there was no statistically significant difference in ADR. The same study also conducted survey among endoscopists regarding their attitudes towards AI-assisted colonoscopy, which demonstrated mixed attitudes towards the new technology. The main concerns were high number of false-positive signals (82.4%), high level of distraction (58.8%), and an impression of prolonged procedure time (47.1%). A pragmatic implementation trial also showed no significant effect of CADe on ADR, APC, or any other detection metric [83]. More studies are required to confirm the ideal role of CADe in real-world practice.

Computer-Aided Characterization

At present, the CADx system has not yet established itself as an independent diagnostic tool to replace the endoscopist. However, we should not undermine its capability just yet. High rates of NPV are achieved especially when there is concordance between the CADx’s output and the optical diagnosis made by expert endoscopists (Table 3). CADx’s complimentary role in this setting provides a greater level of confidence among the expert endoscopists to adopt the “diagnose-and-leave” strategy. This was also affirmed by Kato et al. [84] in their retrospective study. In this study, endoscopists were tasked to characterize still images of polyps into neoplastic versus non-neoplastic. The use of NBI-CAD (with endoscopist) was associated with a significant improvement in the NPV of 85.8% as compared to an NPV of only 75.6% for endoscopist alone. However, more studies are still required to confirm CADx’s complimentary role in the real-world setting.

Discussion

Future Directions

Quality Control

Strict quality indices are crucial in ensuring high-quality screening colonoscopy in prevention of CRC. While CAQ is able to provide real-time feedback to the endoscopists on their performance and quality of colonoscopy, there are a few areas that need to be addressed. For example, as demonstrated earlier in this review paper, the training datasets during the machine learning phase are often smaller in size and limited in variability (single endoscopy systems, small number of endoscopists from which the colonoscopy videos were recorded) compared to CADe and CADx systems. This is crucial in quality control, where the scope movements, withdrawal speeds, and stability of the visual display vary from endoscopist to endoscopist. Developing an optimal CAQ system to standardize such practices with automated and objective feedback to limit such variabilities in colonoscopy quality hence requires more robust datasets for training. This should ideally encompass a large volume of colonoscopy videos from different endoscopists in various centres and with different training experiences. The trained models should also be validated externally with a high procedural volume of colonoscopies to accurately detect any changes in the ADR, which is the most often used endpoint in colonoscopy studies. Lastly, several studies are prospective observational studies or have small sample sizes. More adequately powered RCTs are necessary to validate the various AI models used in CAQ.

Computer-Aided Detection

Currently, there are still many areas of CADe lacking sufficient data and requiring more future studies: (1) the long-term effectiveness on the PCCRC prevention in a longitudinal follow-up; (2) the cost-effectiveness in different geographical areas due to huge differences in health-care systems, socioeconomic situations, and reimbursement policies internationally; (3) how to implement CADe in real-world clinical settings and promote the use of CADe by active engagement of endoscopists, health-care workers, and patients.

Computer-Aided Characterization

Despite its limited applicability in the current clinical climate, CADx is not far from being integrated into our daily practice. Continuous improvements in the DL algorithms are still required to better improve the quality of the output. In fact, having larger collaborations across institutions and countries would allow sharing of the wealth of images, videos, and other data to achieve that. With a large dataset, future studies can capitalize on individual imaging capabilities of white light imaging, blue light imaging, and NBI and harmonize them into one system.

Other Applications

More functions of AI in colonoscopy are being developed and expected to be available in the near future, including real-time polyp size measurement [85, 86]. With an accurate size estimation, one can select the most appropriate polypectomy method and decide further surveillance interval. Moreover, a pre-clinical study demonstrated the possibility of AI-assisted delineation of blood vessels and dissection planes during third-space therapeutic endoscopy, which may potentially reduce intra-procedural complications and shorten learning curves of these complex procedures [87].

Conclusion

This review provided a comprehensive overview of current evidence of AI in colonoscopy. It is clear that AI plays a role in quality improvement, polyp detection, and characterization during colonoscopy, but on the other hand, it also has some drawbacks which may hinder the further implementation in clinical use (Table 5). Long-term data on clinical efficacy, cost-effectiveness, liability, and data sharing are the key areas to be addressed. It is believed that further improvement and new development of AI will bring us to a higher level in colonoscopy field in near future.

Table 5.

Core tips for artificial intelligence in colonoscopy

	Benefits	Limitations
CAQ	Highlighting blind spots Real-time monitoring of withdrawal speed Objective assessment of bowel preparation	AI algorithm does not validate with other endoscopy system Patients with special conditions were excluded from training sets leading to failure of AI algorithm to recognize certain pathology
CADe	Increases ADR, PDR, PPC, APC Reduces AMR, PMR	Does not increase detection rates of advanced or large (>10 mm) adenomas, SSLs, and CRC The false-positive signals may increase the psychological distraction of the endoscopists, leading to a longer inspection and withdrawal time May result in deskilling of endoscopists
CADx	Provides a high level of confidence for expert endoscopists to adopt the “diagnose-and-leave” strategy, especially when there is concordance between the CADx and the optical diagnosis made by them	Limited role as an independent reader SSLs are either under-represented or misrepresented by most studies Long-term impact on cost-effectiveness is unclear

ADR, adenoma detection rate; AMR, adenoma miss rate; APC, adenoma per colonoscopy; CAQ, computer-aided quality improvement; CADe, computer-aided detection; CADx, computer-aided characterization; CRC, colorectal cancer; PDR, polyp detection rate; PMR, polyp miss rate; SSLs, sessile serrated lesions.

Conflict of Interest Statement

LHSL has research collaborations with Olympus Co. Ltd. and GenieBiome Ltd., and served as advisory board for AstraZeneca and as lecture speaker for Olympus Co. Ltd., Boston Scientific Co. Ltd., GenieBiome Ltd., and Pfizer Inc. PWYC has research collaboration with Olympus Co. Ltd. and Boston Scientific, and served as an advisor for EndoVision and EndoMaster and as a lecture speaker for Olympus. All other co-authors do not have any conflicts of interest to disclose.

Funding Sources

This study was not supported by any research grant.

Author Contributions

W.Y. Lai, Kenneth Lin Weicong, and Loi Pooi Ling contributed equally on literature review, data analysis, and manuscript writing. James W. Li, Louis H.S. Lau, and Philip W.Y. Chiu were responsible for conceptualization and critical review of manuscript.

Funding Statement

This study was not supported by any research grant.