Artificial intelligence in colonoscopy: Enhancing quality indicators for optimal patient outcomes

Abstract

Colonoscopy remains the cornerstone of colorectal cancer prevention and surveillance, but the procedure’s effectiveness is entirely dependent upon various quality indicators, such as detection rates, withdrawal time, adequate bowel preparation, cecal intubation rate and patient outcomes. Despite progress in endoscopic techniques, challenges persist in maintaining endoscopists’ consistent performance and improving quality metrics. Artificial intelligence (AI) has emerged as a “game changer” in the gastroenterology field, offering the opportunity to significantly increase colonoscopy quality. This review highlights the role of AI-driven technologies such as deep learning, computer vision, and real-time feedback mechanisms in optimizing key quality indicators of colonoscopy. The implementation of AI in colonoscopy may reduce human error, improve endoscopist’s consistency in real-time decision making, ensuring higher reliability and standardization during the procedure. Furthermore, AI has the potential to reshape how endoscopists perform and evaluate procedures, while improved lesion characterization may enable more precise selection for resection, reducing morbidity and the incidence of interval cancers. The review also addresses challenges and limitations in AI integration, including cost-effectiveness and its impact on endoscopist training. AI holds substantial promise for advancing colonoscopy quality and elevating overall patient care, paving the way for more effective and personalized medical approaches.

Key Words: Artificial intelligence; Colonoscopy; Outcome; Quality indicators; Detection rates

Core Tip: Artificial intelligence (AI) has emerged as a breakthrough innovation in modern colonoscopy, offering significant improvements in key quality indicators, including adenoma detection rate, polyp detection rate, adenoma miss rate, bowel preparation assessment, withdrawal time, and cecal intubation recognition. By optimizing lesion characterization, supporting optical diagnosis and assisting in invasion depth prediction, AI may lead to early diagnosis and prevention of colorectal cancer. This review summarizes the current evidence on the application of AI in key colonoscopy quality indicators, highlighting its role in standardizing practice, improving patient outcomes, and advancing personalized care while also addressing the challenges of training, cost-effectiveness, and ethical considerations.

INTRODUCTION

Colorectal cancer (CRC) remains one of the leading causes of cancer-related morbidity and mortality worldwide, accounting for approximately 10% of all cancer cases[1]. Early detection through population-based CRC screening programs can significantly reduce both incidence and mortality, with colonoscopy considered the gold standard among available modalities, owing to its ability to simultaneously detect and remove pre-malignant lesions[2-4]. Although colonoscopy is highly effective for diagnosis and treatment, it is a resource-intensive and heavily endoscopist-dependent procedure, and it remains far from a perfect solution[5]. However, when performed with high quality, it is considered cost-effective for CRC screening and prevention, leading to a reduction in the incidence of post-colonoscopy CRC (PCCRC) and the related healthcare burden[6].

The clinical efficacy of colonoscopy is directly linked to specific quality indicators, such as adequate bowel preparation, cecal intubation, withdrawal time (WT), adenoma detection rate (ADR), and appropriate lesion characterization and treatment[7-9]. Low-quality colonoscopy has been associated with increased adenoma miss rate (AMR) and a higher PCCRC incidence and mortality[10]. Additionally, the wide variability in endoscopists’ skills and experience, combined with factors such as fatigue and procedural distractions, results in inconsistent outcomes, including elevated rates of missed and inaccurately characterized lesions and inadequate resection[5,11]. Consequently, improving quality metrics, as well as consistency, reliability and generalizability of colonoscopy performance between different endoscopists and clinical settings, is a crucial goal in the rapidly evolving field of modern endoscopy.

In recent years, artificial intelligence (AI) has emerged as a promising tool in gastroenterology, potentially marking the beginning of a new era in next-generation endoscopic approaches and clinical practice. AI includes a range of computational software algorithms, most notably machine learning (ML) which enables systems to learn from data and improve their performance without explicit programming[12]. Deep learning (DL), a more advanced subset of ML based on artificial neural networks, is particularly effective at automatically processing complex and unstructured data, without the need for human intervention[13]. In colonoscopy, these technologies are applied in computer-aided detection (CADe), to identify lesions in real-time, and in computer-aided diagnosis (CADx), where models are trained on large amounts of pre-diagnosed and pre-classified polyp images, aiding optical lesion characterization, increasing diagnostic accuracy, and supporting clinical decision-making[14]. These AI-driven applications are of paramount importance, demonstrating strong potential in improving priority quality indicators, standardizing procedural performance, and reducing inter-operator variability.

The aim of this review is to summarize the current evidence on the application of AI in improving colonoscopy quality indicators, enhancing real-time decision-making, and ultimately contributing to improved patient outcomes. In addition, we address the challenges and limitations of AI implementation, including its cost-effectiveness, and potential impact on training, in order for it to be successfully integrated into clinical practice and support more personalized and standardized endoscopic care.

CURRENT GUIDELINES AND SOCIETY POSITIONS ON AI APPLICATIONS IN COLONOSCOPY

Clear clinical guidance from international committees on the indications, optimal use, limitations, and potential harms of AI tools in colonoscopy is essential. A few formal practice guidelines and position statements on the use of AI in colonoscopy have been recently published. The European Society of Gastrointestinal Endoscopy (ESGE) has released a general position statement establishing the framework for the role of AI in gastrointestinal endoscopy. In general, ESGE suggests that the implementation of AI technology during colonoscopy should aim to standardize diagnostic accuracy by facilitating less experienced endoscopists instead of focusing on marginal gains for experts. AI technology could support the evaluation of bowel preparation adequacy, helping reduce interobserver variability and standardize mucosal cleanliness scoring. It may also assist in evaluating the completeness of the examination by recognizing the cecal landmarks and assessing the successful mucosal exposure. Nevertheless, ESGE underlines that AI-assisted evaluation should present outcomes equivalent to those of skillful endoscopists before being incorporated in clinical practice.

Regarding detection, ESGE supports AI-based CADe systems to decrease AMR, provided false positive signals are minimal and WT is not negatively influenced. Moreover, CADx could aid in the characterization of diminutive (≤ 5 mm) polyps and guide significant clinical decisions with the application of “leave-in-situ” and “resect and discard” strategies, when diagnostic accuracy meets established standards. Finally, AI technology may estimate the invasion depth in larger polyps and determine the feasibility of endoscopic resection (ER) if their performance matches expert assessment[15].

In December 2024, ESGE recommended the use of CADe technology for CRC screening and polyp surveillance. Nevertheless, this recommendation was considered weak due to the lack of evidence on clear benefit, as the limited reduction in CRC incidence and mortality is outweighed by the risk of over diagnosing clinically insignificant lesions and increasing surveillance burden[16]. The consensus was based on data from the BMJ living practice guideline developed by the Making GRADE the Irresistible Choice Evidence Ecosystem Foundation and the BMJ Rapid Recommendation series[17]. Interestingly, although the BMJ guideline was based on the same systematic review and meta-analysis of 44 randomized trials assessing CADe performance, it recommended against the routine use of CADe, citing concerns about unnecessary intensive surveillance and potential psychological burden for patients. This recommendation was also considered “weak” due to uncertainty regarding critical outcomes such as CRC incidence, mortality or PCCRC[18].

Similarly, the American Gastroenterological Association issued a guideline on CADe colonoscopy, in collaboration with the BMJ panel based on the same data, but finally refrained from making a definitive suggestion for or against CADe use[19]. Consequently, each gastroenterology society appears to evaluate and weigh patient preferences and cost-effectiveness of AI technology application during colonoscopy differently, leading to conflicting recommendations.

The World Endoscopy Organization also acknowledges CADe’s potential to maximize colonoscopy efficiency and polyp detection. However, it notes concerns about possible short-term healthcare cost increase due to AI technology expenses, more frequent polypectomies and histopathological analyses. Over time, these costs might be compensated by CRC incidence reduction. The statement encourages national health systems and authorities to assess the cost-effectiveness of CADe and CADx systems and highlights the need for high-quality studies across various healthcare systems in order to appropriately guide patient care[20]. To support practical implementation, Table 1 summarizes the key take-home messages for clinicians.

Table 1 Take-home messages for practicing endoscopists based on current guidelines.

Society	Current position on AI in colonoscopy
ESGE[16]	Supports CADe use to reduce AMR, provided minimal false positives and no WT prolongation; allows CADx for “leave-in-situ”/“resect and discard”, if accuracy is equivalent to experts. Recommends AI integration to help standardize performance in less experienced endoscopists
BMJ[17]	Recommends against routine CADe use due to concerns over overdiagnosis, minimal clinical benefit, and increased surveillance; recommendation graded as “weak”
AGA[19]	Does not issue a definitive recommendation for or against CADe use, citing uncertainty in long-term outcomes
WEO[20]	Acknowledges CADe/CADx potential but urges cost-effectiveness assessment; highlights the need for high-quality real-world studies across healthcare systems

AI: Artificial intelligence; ESGE: European Society of Gastrointestinal Endoscopy; AGA: American Gastroenterological Association; WEO: World Endoscopy Organization; CADe: Computer-aided detection; AMR: Adenoma miss rate; CADx: Computer-aided diagnosis; WT: Withdrawal time.

AI-ASSISTED COLONOSCOPY AND KEY QUALITY INDICATORS

AI-assisted bowel preparation: Evaluation and optimization

Strategies to evaluate bowel preparation scores based on computer automated algorithms have begun to emerge since the last decade. In a simplistic overview most of these frameworks are built using DL convolutional neural networks (DCNNs) trained on large datasets of human-annotated images and videos, and are validated by comparing machine-generated values with human-derived metrics. These technologies range from early image analysis tools to advanced real-time AI systems and patient-friendly smartphone applications. Overall, they aim to standardize bowel preparation evaluation, improve detection rates, and optimize both endoscopist performance and patient compliance.

Early image-based scoring systems: The clean colon software program (CCSP) represented an early effort to automate bowel cleanliness scoring[21]. Researchers employed image recognition algorithms using color analysis to differentiate clean from dirty mucosa and to compute a continuous cleanliness score. Software derived scores were compared to endoscopist Boston bowel preparation system (BBPS) ratings, resulting in high interclass correlation [up to 0.87, 95% confidence interval (CI): 0.84-0.90]. However, CCSP lacked real-time application and struggled under challenging visual conditions, limiting its clinical utility[21].

Real-time AI systems: The automatic quality control system (AQCS) combined multiple AI modules to assess WT, mucosal stability, bowel cleanliness, and polyp detection[22]. AQCS significantly improved ADR, bowel preparation adequacy, and compliance with recommended WT, using real-time prompts including audio alerts[22]. A more refined system, ENDOANGEL, achieved 93.33% accuracy in bowel preparation assessment, outperforming human raters[23]. Its bowel preparation module, e-BBPS, was validated in 616 patients, predicting ADR ≥ 25% for scores ≤ 3 and identifying suboptimal preparation linked to higher AMR (35.71% vs 13.19%, P = 0.0056)[24,25].

Lee et al[26] and Lee et al[27] developed a model trained on still images and validated on videos, achieving 100% sensitivity for inadequate preparation, though clinical impact remains untested. Similarly, Feng et al[28] introduced a novel three dimensional-CNN based AI system using full-length videos[28], achieving high accuracy [area under the receiver operating characteristic curve (AUROC) = 0.98, 95.2%] in characterizing bowel preparation as adequate (BBPS 2-3) or inadequate (BBPS 0-1)[28]. However, it has not yet surpassed endoscopist accuracy or been clinically applied. The open-source automatic bowel preparation scale, a non-BBPS-based model, computes a continuous score using mucosa-to-fecal pixel ratios[29]. It showed moderate-to-strong negative correlations with BBPS scores (r = -0.70) and was associated with polyp detection rates (PDR)[29]. As an open-source platform, open-source automatic bowel preparation scale invites further development and evaluation across diverse clinical settings.

Smartphone based applications: An emerging application of AI in colonoscopy lies in optimizing bowel preparation before the procedure. Several studies have developed DCNN models to evaluate stool images taken during the purgative phase, integrating them into smartphone applications for patient use and providing real-time feedback. If stool quality does not meet pre-established thresholds, patients receive tailored recommendations to receive additional purgatives or enemas. Additional app features, such as reminder alerts and visual guidance, further support adherence to preparation protocols[30-32]. Inaba et al[30] reported over 99.0% (95%CI: 94.8%-100%) adequate preparation (BBPS ≥ 6) with ADR, cecal intubation rate, and WT meeting or exceeding ESGE standards. Patient feedback was unanimously positive.

Two studies comparing AI apps to standard educational materials exhibited significantly better BBPS scores, shorter cecal intubation times and higher PDR in the AI groups[31,32]. Patients using the apps demonstrated higher adherence, better sleep quality and willingness to reuse and recommend the app[31,32]. However, the mainly East Asian study population hinders generalizability to other healthcare settings with different dietary patterns and patient demographics. Moreover, smartphone-based application necessity may exclude senior patients or patients with limited access to such devices.

In summary, AI-based approaches for evaluating and optimizing bowel preparation are very promising to standardize cleanliness assessment and improve colonoscopy quality. Continued validation in diverse populations and settings is crucial to enhance their effectiveness in routine practice.

AI in cecal intubation recognition

Cecal intubation is an important quality indicator for colonoscopy, with various guidelines recommending an acceptable rate of > 95%[7,33]. Therefore, cecal intubation has been incorporated in various AI models, either as part of an integrated model or as a standalone function.

Low et al[34] created a standalone DCNN based model that predicts cecal intubation based on appendiceal orifice identification with high sensitivity and specificity of 96% and 92% respectively. Accuracy remained > 95% across varying bowel preparation quality. Another AI cecal intubation recognition system compared clinical metrics before and after system implementation in local clinical practice[35]. ADR and advanced ADR significantly increased, especially in the proximal colon but overall cecal intubation rates did not differ. Thorough cleaning and careful inspection of the proximal colon were essential to meet verification criteria, potentially contributing to the observed increase in proximal ADR[35].

AI-assisted assessment of withdrawal speed in colonoscopy

Withdrawal speed during colonoscopy has long been considered a surrogate marker of procedural quality, grounded in evidence linking prolonged WT with increased ADR. While current guidelines recommend a minimum WT of 6 minutes, recently increased to 8 minutes in the latest American Society for Gastrointestinal Endoscopy guidelines emerging data suggest that withdrawal speed alone is an insufficient marker of mucosal inspection quality[7,33]. This has spiked interest in AI-based systems that offer real-time, frame-by-frame assessment of withdrawal phase.

Gong et al[36] introduced the proportion of overspeed frames (POF), a novel AI-derived metric capturing frames with withdrawal speed ≥ 44. In 1804 procedures, POF demonstrated strong inverse correlation with ADR (r = -0.836). Even with WT ≥ 6 minutes, ADR was higher when POF was ≤ 10%, suggesting WT alone may be insufficient. Building on the theme of speed monitoring, Barua et al[37] evaluated a real-time speedometer system with an acoustic alarm for over speeding. Despite its intuitive design, the intervention did not significantly alter mean WT or ADR compared to standard practice, indicating limited impact of speed-only feedback. A more integrative approach was proposed by Filip et al[38] through the Colometer system, which combined withdrawal speed, image clarity, and stool coverage to compute an overall quality score. In this pilot study, system’s automated ratings showed moderate correlation with expert endoscopists’ assessments, demonstrating feasibility for real-time feedback during colonoscopy.

These studies highlight a key conceptual shift: Rather than measuring time or speed in isolation, AI tools offer the capacity to assess various withdrawal dynamics as they relate to visual stability and mucosal coverage. This is demonstrated by a recent study that proposed effective WT (EWT) a time-weighted metric derived from AI classification of visual fields throughout withdrawal[39]. EWT correlated more strongly with ADR than standard WT (AUROC: 0.80 vs 0.70, P < 0.01), with each minute increase associated with 49% higher odds of adenoma detection. Unlike speed alone, EWT accounts for the quality of mucosal visualization, offering a more robust and clinically relevant marker of inspection quality.

In summary, while withdrawal speed remains an important procedural parameter, AI-enhanced approaches now enable a more nuanced assessment of inspection behavior. Metrics such as POF and EWT bridge the gap between raw time and effective visualization, positioning AI as a critical tool for procedural audit, training, and quality assurance in colonoscopy.

AI-assisted colonoscopy: Effects on detection and miss rates

International gastroenterology societies recommend ADR as a key quality indicator during colonoscopy for ensuring adequate inspection of intestinal mucosa[7,33]. A minimum ADR of approximately 25% overall, 30% for men and 20% for women, is suggested for screening colonoscopy[33]. AI-based CADe systems may significantly improve ADR by detecting lesions often missed due to subtle morphology, difficult location, suboptimal bowel preparation, or endoscopist fatigue.

ADR: Over the past decade, numerous studies and meta-analyses have consistently demonstrated that CADe tools significantly improve ADR during colonoscopy (Tables 2 and 3). A recent meta-analysis of 44 randomized clinical trials (RCTs) revealed an improved ADR with CADe systems compared with conventional colonoscopy [44.7% vs 36.7%; relative risk (RR) = 1.21 (95%CI: 1.15-1.28)][18]. Similarly, Makar et al[40] reported a 20% increase in ADR with AI technology in a meta-analysis of 28 prospective RCTs [RR = 1.20 (95%CI: 1.14-1.27)].

Table 2 Summary of systematic reviews and meta-analyses evaluating the impact of artificial intelligence on detection rate metrics in colonoscopy[18,37,40,41,43,44,46,48,50,51,55-61,108-113].

Ref.	Year	Study type	Patient number	Number of studies	Primary outcome	Artificial intelligence vs conventional colonoscopy (95%CI)
Barua et al[37]	2020	Systematic review and meta-analysis	4311	5 RCTs	ADR, PDR, APC, PPC, aAPC	ADR: 29.6% vs 19.3%; RR = 1.52 (1.31 to 1.77); PDR: 45.4% vs 30.6%; RR = 1.48 (1.37 to 1.60); No difference in aAPC; PPC: 0.93 vs 0.51; Mean difference 0.42 (0.33 to 0.50); APC: 0.41 vs 0.23; Mean difference 0.18 (0.13 to 0.22)
Aziz et al[108]	2020	Systematic review and meta-analysis	2815	3 RCTs	ADR	32.9% vs 20.8%; RR = 1.58 (1.39 to 1.80)
Mohan et al[109]	2020	Systematic review and meta-analysis	4962	6 RCTs	ADR	32.8% vs 21.1%; RR = 1.5 (1.30 to 1.72)
Ashat et al[110]	2021	Systematic review and meta-analysis	5058	6 RCTs	ADR	33.7% vs 22.9%; OR = 1.76 (1.55 to 2.00)
Hassan et al[56]	2021	Systematic review and meta-analysis	4354	5 RCTs	ADR	36.6% vs 25.2%; RR = 1.44 (1.27 to 1.62)
Deliwala et al[61]	2021	Systematic review and meta-analysis	4996	6 RCTs	ADR, PDR	ADR: OR = 1.77 (1.50 to 2.08); PDR: OR = 1.91 (1.68 to 2.16)
Nazarian et al[51]	2021	Systematic review and meta-analysis	5577 subjects for polyp detection with ADR and PDR	48 studies (18 studies for polyp detection, 22 studies for polyp characterization and 8 studies for PDR)	ADR, PDR, polyp characterization	PDR: OR = 1.75 (1.56 to 1.96); ADR: OR = 1.53 (1.32 to 1.77)
Li et al[59]	2021	Systematic review and meta-analysis	4311	5 studies	ADR, PDR	PDR: OR = 1.91 (1.68 to 2.16); ADR: OR = 1.75 (1.52 to 2.01)
Zhang et al[60]	2021	Systematic review and meta-analysis	5427	7 RCTs	ADR, PDR	PDR: OR = 1.95 (1.75 to 2.19); ADR: OR = 1.72 (1.52 to 1.95)
Spadaccini et al[48]	2021	Systematic review and network meta-analysis	34445	50 RCTs	ADR	CADe vs HD white-light endoscopy OR = 1.78 (1.44 to 2.18); CADe vs chromoendoscopy OR = 1.45 (1.14 to 1.85); CADe vs increased mucosal visualization systems OR = 1.54 (1.22 to 1.94)
Huang et al[41]	2022	Systematic review and meta-analysis	6629	10 RCTs	ADR, PDR	ADR: 35.4% vs 24.9%; RR = 1.43 (1.33 to 1.53)/OR = 1.45 (1.32 to 1.59); PDR: 48.6% vs 33.8%; RR = 1.44 (1.35 to 1.53)/OR = 1.90 (1.70 to 2.11)
Shah et al[62]	2023	Systematic review and meta-analysis	10928	14 RCTs	ADR, PDR	ADR: OR = 1.52 (1.39 to 1.67); PDR: OR = 1.48 (1.37 to 1.61)
Hassan et al[57]	2023	Systematic review and meta-analysis	18232	21 RCTs	ADR, APC, aAPC, number of serrated lesions/colonoscopy, number of polypectomies for nonneoplastic lesions, WT	ADR: 44.0% vs 35.9%; RR = 1.24 (1.16 to 1.33)
Lou et al[58]	2023	Systematic review and meta-analysis	27404	33 RCTs	ADR, APC	ADR: RR = 1.242 (1.159 to 1.332); APC: IRR = 1.390 (1.277 to 1.513)
Adiwinata et al[50]	2023	Systematic review and meta-analysis	NA	13 RCTs	ADR, PDR	PDR: OR = 1.46 (1.13 to 1.89); ADR: OR = 1.58 (1.37 to 1.82)
Shiha et al[43]	2023	Systematic review and meta-analysis	11340	12 RCTs	ADR	41.4% vs 33%; RR = 1.26 (1.18 to 1.35)
Wei et al[111]	2024	Systematic review and meta-analysis	11660	12 studies	ADR	ADR was statistically significantly higher with AI vs CC (36.3% vs 35.8%, RR = 1.13, 95%CI: 1.01 to 1.28)
Aziz et al[112]	2024	Systematic review and network meta-analysis	61172	94 RCTs	ADR	Autofluorescence imaging: RR = 1.33 (1.06 to 1.66); Dye-based chromoendoscopy: RR = 1.32 (1.17 to 1.50); Endo cuff: RR = 1.19 (1.04 to 1.35); Endo cuff vision: RR = 1.26 (1.13 to 1.41); Endoring: RR = 1.30 (1.10 to 1.52); Flexible spectral imaging color enhancement: RR = 1.26 (1.09 to 1.46); Full-spectrum endoscopy: RR = 1.40 (1.19 to 1.65); High definition: RR = 1.41 (1.28 to 1.54); Linked color imaging: RR = 1.21 (1.08 to 1.36); Narrow band imaging: RR = 1.33 (1.18 to 1.48); Water exchange: RR = 1.22 (1.06 to 1.42); Water immersion: RR = 1.47 (1.19 to 1.82)
Soleymanjahi et al[18]	2024	Systematic review and meta-analysis	36201	44 RCTs	APC, ACN	APC: 0.98 vs 0.78; IRD = 0.22 (0.16 to 0.28); ACN: 0.16 vs 0.15; IRD = 0.01 (-0.01 to 0.02)
Patel et al[46]	2024	Systematic review and meta-analysis	9782	8 non-randomized studies	ADR	44% vs 38%; RR = 1.11 (0.97 to 1.28)
Gangwani et al[113]	2024	Network meta-analysis	22560	26 studies (20 RCTs, 3 retrospective, 3 prospective studies)	ADR	AI vs single operator: RR = 1.1 (0.9 to 1.2)
Lee et al[44]	2024	Systematic review and meta-analysis	17413	24 RCTs	ADR	Tandem studies: RR = 1.18 (1.08 to 1.30); Parallel studies: RR = 1.26 (1.17 to 1.35); Overall ADR: RR = 1.24 (1.17 to 1.31)
Makar et al[40]	2025	Systematic review and meta-analysis	23861	28 RCTs	ADR	ADR: RR = 1.20 (1.14 to 1.27)
Spadaccini et al[55]	2025	Systematic review and meta-analysis	5421	10 RCTs	ADR	0.62 vs 0.52; RR = 1.19 (1.08 to 1.31)

CI: Confidence interval; RCTs: Randomized clinical trials; ADR: Adenoma detection rate; PDR: Polyp detection rate; NR: Not reported; APC: Adenomas per colonoscopy; PPC: Mean number of polyps per colonoscopy; aAPC: Mean number of advanced adenomas detected per colonoscopy; WT: Withdrawal time; ACN: Advanced colorectal neoplasia; RR: Relative risk; OR: Odds ratio; HD: High definition; CADe: Computer-aided detection; AI: Artificial intelligence; IRR: Incidence rate ratio; CC: Conventional colonoscopy; IRD: Incidence rate difference.

Table 3 Summary of systematic reviews and meta-analyses evaluating the impact of artificial intelligence on miss rate metrics in colonoscopy[40,52,57,58].

Ref.	Year	Study type	Patient number	Number of studies	Primary outcome	Artificial intelligence vs conventional colonoscopy (95%CI)
Hassan et al[57]	2023	Systematic review and meta-analysis	18232	21 RCTs	AMR	AMR: 16% vs 35%; RR = 0.45 (0.35 to 0.58)
Lou et al[58]	2023	Systematic review and meta-analysis	27404	33 RCTs	AMR	AMR: RR = 0.495 (0.390 to 0.627)
Maida et al[52]	2024	Systematic review and meta-analysis	1718	6 RCTs	AMR, PMR	PMR: 16.3% vs 38.1%; RR = 0.44 (0.33 to 0.60); AMR: 15.3% vs 34.1%; RR = 0.46 (0.38 to 0.55)
Makar et al[40]	2025	Systematic review and meta-analysis	23861	28 RCTs	AMR	AMR: RR = 0.45 (0.37 to 0.54)

CI: Confidence interval; RCTs: Randomized clinical trials; AMR: Adenoma miss rate; PMR: Polyp miss rate; RR: Relative risk.

Despite this general benefit, considerable variation exists among studies due to methodological, clinical, and technological factors[41]. Variability in study colonoscopy indication [screening vs surveillance endoscopy vs fecal immunochemical test (FIT) positive], polyp prevalence and bowel preparation adequacy all affect ADR outcomes[41,42]. Single center studies and those conducted in Asia often report a higher ADR gain compared to multicenter trials and studies from Europe or the United States[43]. Another major issue is that multiple meta-analyses often include heterogeneous AI systems with varying designs and performance characteristics, limiting the ability to draw consistent conclusions. Future research should include head-to-head trials directly comparing different commercially available or research-based AI tools. Evidence indicates that AI increases ADR regardless of endoscopic experience; even experienced endoscopists can increase their ADR by an average of 19%[40,44,45]. Nevertheless, further research should emphasize how CADe affects baseline ADR, rather than operator experience level alone.

Study design also influences outcomes; a meta-analysis demonstrated both tandem and parallel studies benefit from AI, though tandem studies showed more prominent ADR improvement[44]. In contrast, a meta-analysis by Patel et al[46], which included non-randomized studies, argued that the lack of blinding in RCTs due to the nature of the intervention, and potential endoscopist overreliance may limit AI’s advantage, reporting no significant ADR [44% vs 38%; RR = 1.11; (95%CI: 0.97-1.28)] or adenomas per colonoscopy (APC) gain. Consequently, although CADe generally improves ADR, more studies evaluating the determinants for optimal ADR increase with AI are crucial.

A recent network meta-analysis reported that AI assisted colonoscopy significantly outperformed established interventions, such as autofluorescence imaging, dye-based chromoendoscopy (CE), distal attachment devices, full-spectrum endoscopy, narrow band imaging (NBI), water exchange and immersion, in ADR improvement[47]. Similarly, Spadaccini et al[48] indicated that CADe was superior to white-light endoscopy (WLE), CE and other mucosal visualization systems for ADR. Future studies examining the combined use of AI with mucosal enhancement techniques are essential to optimize colorectal lesions detection.

PDR: PDR is defined as the proportion of colonoscopies in patients over 50 years old in which at least one polyp is detected, serving as a practical alternative to ADR since no histological confirmation is required and it can be calculated in real-time[49]. AI implementation during colonoscopy has shown promising results in raising PDR compared to conventional colonoscopy with pooled PDRs of approximately 48%-49% using CADe vs 33%-34% with standard colonoscopy (RR = 1.44; 95%CI: 1.35-1.53), and reported ranges from 38% to 64%[50,51]. Although several meta-analyses have shown that the AI systems improve PDR compared to standard colonoscopy (Figure 1), Patel et al[46] found no differences among AI-assisted and conventional colonoscopy based on data exclusively from non-randomized studies.

Figure 1 Summary of systematic reviews and meta-analyses assessing artificial intelligence impact on colonoscopy quality metrics[18,37,40,41,43,44,46,48,50-52,57-62,89,108-112]. ¹Advanced adenomas were defined according to current guidelines as size ≥ 10 mm, and/or with villous components > 20%, and/or high-grade dysplasia. CADe: Computer-aided detection; CC: Convetional colonoscopy.

APC: The mean number of APC is considered as a useful quality indicator during colonoscopy and it is strongly correlated with ADR. Meta-analyses consistently show that AI-assisted colonoscopy significantly increases the mean APC compared to conventional colonoscopy. In a recent meta-analysis the average APC with CADe was 0.98 compared to 0.78 for standard colonoscopy [incidence rate difference = 0.22 (95%CI: 0.16-0.28)][18].

AMR: AMR, defined as the proportion of adenomas missed during initial colonoscopy to the total adenomas found during sequent colonoscopy, is also a key performance indicator, as missed polyps contribute to interval CRC risk. In a meta-analysis of 15000 tandem colonoscopies, AMR was estimated 26% for adenomas, 9% for advanced adenomas and 27% for serrated polyps[5]. Studies investigating the role of CADe on AMR are more challenging due to the requirement of tandem colonoscopies, and recent evidence demonstrates that CADe significantly decreases AMR from 34.1% to 15.3% (RR = 0.46; 95%CI: 0.38-0.55), with reductions of up to 55%[40,52]. Despite this, CADe does not consistently outperform conventional colonoscopy in detecting advanced adenomas (Figure 1). This could be explained by the fact that some advanced features, such as villous histology or high-grade dysplasia, are confirmed after histological analysis, and are not visually identified. Larger adenomas (> 10 mm) can easily be detected by endoscopists when mucosal exposure is optimal, while specific types such as lateral spreading tumors may be undistinguishable from normal mucosa and are not specifically targeted by CADe algorithms. In this context, Soleymanjahi et al[18] concluded that there was a small, clinically insignificant increase in advanced colorectal neoplasia detection rate with CADe (12.7% vs 11.5%, RR = 1.16, 95%CI: 1.02-1.32).

Sessile serrated lesion detection rate: Another novel quality indicator is the sessile serrated lesion detection rate (SSLDR), defined as the proportion of patients over 45 undergoing colonoscopy with ≥ 1 sessile serrated lesions (SSLs) removed and confirmed by pathology[7]. SSLs constitute preneoplastic lesions which play crucial role in the serrated pathway of colorectal carcinogenesis, accounting for approximately 20% of CRC cases, particularly interval CRC[53]. Their subtle, mucus-capped, cloud-like appearance, indistinct borders, and frequent location in the proximal colon make them challenging to detect, especially with inadequate bowel preparation[54]. These features lead to high interobserver variability, even among experienced endoscopists, and limit the effectiveness of traditional detection techniques. Although AI could theoretically enhance detection, CADe systems are not usually designed to identify SSL. Most CADe tools have been trained mainly on conventional adenomas, resulting in suboptimal performance when applied to SSLs. Results from RCTs and meta-analyses concerning the improvement of SSLDR with CADe systems are conflicting (Figure 1). Huang et al[41] revealed higher SSLDR with AI, while Spadaccini et al[55] reported improved detection in FIT-positive patients. On the other hand, other studies reported no significant effect or only limited benefits with high heterogeneity[40,43]. These findings emphasize the need for CADe systems trained specifically to recognize the challenging visual features of SSLs. Therefore, more prospective randomized trials are needed, focusing exclusively on SSL detection by CADe systems using AI-algorithms specifically trained to identify these lesions. Such algorithms may enhance sensitivity and reduce variability in SSL detection supporting early diagnosis and reducing interval CRC incidence.

Polyp morphology: Regarding morphological characteristics, AI improves detection of both flat and polypoid lesions[56-58]. Although AI is more efficient in detecting flat and sessile adenomas, standard colonoscopy performs equally or better than AI in detecting pedunculated polyps, possibly due to their high visual detectability and variable stalk morphology, which may pose challenges for AI-based recognition algorithms[41,59,60]. AI enhances detection of diminutive adenomas (≤ 5 mm), however, its benefit for polyps sized 5 mm-10 mm and > 10 mm is mixed (Figure 1)[40,43,57]. Whether detecting and removing these lesions reduces CRC risk remains an ongoing discussion due to concerns for increased procedure time, cost and surveillance. AI-assisted colonoscopy improves lesion detection throughout the colon; yet meta-analyses report limited improvement in the ascending colon and cecum, likely due to the shape of colonic folds, and inadequate bowel preparation[41,43,57-59,61,62]. To sum up, it appears that morphological features of polyps such as shape, size and location affect CADe system effectiveness in colorectal lesion detection, highlighting the need for further research to train and validate AI tools for diverse lesion types.

AI-assisted characterization and risk stratification of colorectal lesions

Accurate therapeutic decisions and complete ER of neoplastic lesions represent crucial quality indicators in colonoscopy[7,33]. Endoscopic en bloc R0 resection may be curative for colorectal lesions with low or high-grade dysplasia, intramucosal carcinoma (Tis), or superficial submucosal invasion (T1a < 1000 μm), provided unfavorable histologic features are absent[63,64]. In contrast, deeper submucosal invasion (T1b ≥ 1 000 μm) is associated with a higher risk of lymph node metastasis and recurrence after ER[65]. Therefore, optical diagnosis and prediction of invasion depth are critical in guiding appropriate therapeutic decisions.

Various CADx models using either WLE or image-enhanced endoscopy (IEE) have been developed to facilitate polyp characterization, differentiation of high-risk adenomas and prediction of invasion depth (Table 4). These systems may optimize management by preventing overtreatment of Tis/T1a CRCs and undertreatment of deeply invasive lesions, thereby avoiding unnecessary surgery, poor outcomes, and increased healthcare costs.

Table 4 Summary of study characteristics evaluating artificial intelligence performance for invasion depth prediction and polyp characterization in colonoscopy.

Ref.	Year	Study type	Image type	AI algorithm	Patient number training set	Patient number validation set	Patient number testing set	Primary outcome	Sensitivity (%)	Specificity (%)	Accuracy (%)
Luo et al[69]	2021	Single center retrospective	WLI	Deep learning	556		137	Invasion depth Tis/T1a vs T1b/> T2	91.2	91	91.1
Minami et al[70]	2022	Single center retrospective	WLI, NBI, CCE	Deep learning	91	49	56	Submucosal invasion depth SM1 vs SM2/3	87.2	35.7	74.4
Lu et al[68]	2022	Multicenter retrospective	WLI, NBI, BLI	Deep learning	305		140	Invasion depth LGD/HGD/IM/SM1 vs SM2/advanced CRC	90.0	94.2	93.8
Nemoto et al[72]	2023	Multicenter retrospective	WLI	Deep learning	1084		400	Invasion depth Tis/T1a vs T1b	59.8	94.4	87.3
Tokunaga et al[73]1	2021	Single center retrospective	WLI	Deep learning	824		211	Invasion depth LGD/HGD/SM1 vs SM2/advanced CRC	96.7	75.0	90.3
Nakajima et al[71]	2022	Multicenter retrospective	WLI	Deep learning	313		44	Invasion depth Tis/T1a vs T1b	81.0	87.0	84.0
Song et al[75]	2020	Single center retrospective	NBI	Deep learning	624		545	Invasion depth SSP/BA/SM1 vs SM2/3	58.8	93.3	81.3
Lui et al[74]	2019	Single center retrospective	WLI, NBI	Deep learning	1652		76	Invasion depth polyps ≥ 2 cm adenoma/SM1 vs SM2	94.6	92.3	94.3
Yao et al[76]	2023	Multicenter retrospective	WLI, IEE	Deep learning	339		198	Invasion depth large SSPs ≥ 10 mm	78.8	96.2	90.4
Racz et al[66]	2022	Single center retrospective	NBI	Machine learning	279			Polyp characterization non-neoplastic vs neoplastic2	92.2	77.6	86.6
Ham et al[67]	2025	Single center retrospective	WLI	Deep learning	2696		476	Polyp characterization low vs high-risk adenomas ≤ 10 mm3	75.6	95.7	93.8

¹Clinical impact data were only explicitly reported in the study, which demonstrated a 24% reduction in unnecessary resections. Other included studies did not provide quantifiable real-world outcome metrics.

²Non-neoplastic polyps were defined as hyperplastic polyps, whereas neoplastic polyps included tubular adenomas, villous adenomas, and sessile serrated lesions.

³High-risk adenomas were defined as high-grade dysplasia or adenomas with villous histology.

AI: Artificial intelligence; WLI: White light imaging; NBI: Narrow band imaging; CCE: Conventional chromoendoscopy; BLI: Blue light imaging; IEE: Image enhanced endoscopy; SM: Submucosal invasion; LGD: Low-grade dysplasia; HGD: High-grade dysplasia; IM: Intramucosal carcinoma; CRC: Colorectal cancer; SSP: Sessile serrated polyp; BA: Benign conventional adenoma.

Regarding polyp characterization, an AI-based polyp histology prediction (AIPHP) software was developed to evaluate NBI colonoscopy images[66]. The overall accuracy of AIPHP was 86.6%, higher in the non-diminutive polyps (92.2%, P = 0.0032), possibly due to mucus covering diminutive hyperplastic polyps interfering with vascular pattern analysis. Another CADx model differentiated low- and high-risk adenomas in polyps ≤ 10 mm using standard WLI[67], showing specificity of 95.7% (95%CI: 93.4%-97.3%), comparable to experts and superior to trainees.

Studies assessing invasion depth have used either WLE, NBI, or a combination of both imaging modalities (Table 4)[68-76]. Sensitivity ranged from 58.8% to 96.7%, specificity from 35.7% to 96.2% and accuracy from 74.4% to 94.3%. While low sensitivity remains a concern, high specificity in identifying lesions suitable for ER is considered more crucial to avoid overtreatment with unnecessary surgical intervention.

Nemoto et al[72] reported approximately 95% specificity of a CADx system for diagnosing T1b CRCs, with performance equivalent to experts and superior to trainees. Similarly, a newly developed CAD system distinguished treatable from untreatable lesions with 96.7% sensitivity, 75.0% specificity, and 90.3% accuracy[73]. Importantly, this diagnostic performance was associated with a 24% reduction in unnecessary polypectomies, highlighting the potential real-world clinical impact of AI-assisted optical biopsy strategies[73].

Endo-CRC, incorporating both WLE and IEE images, achieved 88.1% accuracy in distinguishing deeply invasive CRC from superficially submucosal invasive lesions, outperforming models using a single modality[68]. An AI-assisted image classifier predicted curative ER feasibility in large (≥ 2 cm) colonic lesions with 85.5% accuracy overall, better with NBI (94.3%) than WLE (76.0%)[74]. Finally, Yao et al[76] introduced a DL based system for large sessile polyps, achieving 90.4% accuracy comparable to experts and superior to all other participants.

At present CADx systems have not yet been established as independent diagnostic tools to replace endoscopists. Nevertheless, they may serve as an essential auxiliary tool, particularly for trainees, to increase diagnostic accuracy and consistency. Despite advancements, differentiating between T1a and T1b CRCs remains challenging, possibly due to the exclusive use of static image training datasets in most studies. Future studies need to focus on real-time, high-definition video training datasets to improve the diagnostic value of CADx systems, contribute to the accurate identification of endoscopically treatable lesions, reduce inter-operator variability, and ultimately enhance the overall quality of colonoscopy.

IMPACT OF AI ON ENDOSCOPY TRAINING

AI technologies have the potential to revolutionize endoscopist training by enhancing learning efficiency, improving diagnostic accuracy, and refining procedural skills. Research in this area is growing, with many studies now recorded in the most relevant clinical trials databases. Several randomized trials have shown that experienced and non-expert endoscopists enhance ADR and APC[50,77,78]. However, many questions remain unanswered. Where in the training process should AI be introduced? Is it beneficial to early trainees/novices, or is it a major distraction? Furthermore, how does it affect the quality of procedures for trainees and novices? Despite multiple published evaluations and societal comments on AI and colonoscopy, data on the impact of AI-based education on trainees are scarce[79-81].

The fact that inexperienced endoscopists detect fewer adenomas than experts presents a challenge for lower gastrointestinal endoscopy[5,82]. A recent video-based study assessed whether AI improves polyp detection[83]. Thirty-three inexperienced endoscopists (> 400 colonoscopies) evaluated videos twice, with and without AI. Sensitivity increased from 86.3% (95%CI: 85.1%-87.5%) to 91.7% (95%CI: 90.7%-92.6%) with AI assistance (P < 0001) and for neoplastic from 85.4% (95%CI: 84.0%-86.7%) to 92.1% (95%CI: 91.1%-93.1%) (P < 0001)[83].

Another concern is that the efficacy and safety of colonoscopies conducted by amateurs using AI are uncertain. A multicenter, randomized, non-inferiority tandem study compared lesion detection among novices, AI-assisted novices, and experts[84]. Participants had a repeat colonoscopy performed by an AI-assisted expert to assess the lesion miss rate. The AMR and polyp miss rates were lower in non-AI novices than in AI-assisted novices [18.82% vs 43.69% (P < 0.001) and 21.23% vs 35.38% (P < 0.001), respectively]. Notably, it has been demonstrated that AI-assisted novice endoscopists achieved non-inferior AMR compared to experienced endoscopists, highlighting the potential of AI to equalize performance and improve training effectiveness[84].

On the other hand, some studies suggest that experts may benefit less from AI[85,86]. An Italian post-hoc analysis of two RCTs using CADe found no significant difference in ADR based on expertise (> 2000 colonoscopies)[77]. Similarly, a Spanish RCT divided endoscopists by ADR (> 40%) and revealed no significant difference in ADR when CADe was used[87,88].

Concerns exist that AI assistance could limit skill development in trainees, especially if it replaces rather than supplements active learning. If CADe is used as a “concurrent read”, it may improve detection[87]. However, if the endoscopist delegates complete responsibility for detection to the CADe, there is a risk of deskilling. Hassan et al[89] demonstrated that CADe recognized polyps faster than the human eye but with a significant false positive rate, potentially enhancing endoscopist dependence and resulting in loss of skills, particularly among non-experts. A recent study using colonoscopy recordings with and without CADe found that it reduced eye movement distance and endoscopists incorrectly identify more polyps, independent of their level of competence[90,91].

To date there is no strong evidence about the impact of AI incorporation on the training path of trainees/novices. It seems that they may benefit from lesion detection, lowering the AMR. Developing appropriate AI educational programs will enhance learning efficiency and diminish the potential risk of deskilling trainees and young endoscopists. A summary of practical recommendations for integrating AI into endoscopy training is provided in Table 5.

Table 5 Practical tips for incorporating artificial intelligence into endoscopy training.

Practical tips for trainees using AI in colonoscopy
Use AI as a complement, not a replacement for clinical judgment
Review false positive alerts to learn distinguishing features
Combine AI with feedback from supervisors
Practice with and without AI assistance
Incorporate AI into video-based self-review
Be aware of potential deskilling with overuse
Participate in structured training programs including AI modules

AI: Artificial intelligence.

DRAWBACKS AND LIMITATIONS OF AI IN COLONOSCOPY AND QUALITY CONTROL

Despite the promising role of AI in enhancing colonoscopy quality, particularly key quality indicators ADR and PDR, its widespread clinical implementation presents several challenges that must be addressed before it can become routine in endoscopic practice.

One of the main limitations is the improvement and training of AI systems themselves. Generalized and standardized training requires access to large amounts of images obtained under real-life conditions, including both high-quality and suboptimal colonoscopy procedures. However, most studies assessing AI performance in colonoscopy are underpowered, involve inadequate or non-representative datasets, and often lack validated training protocols, raising concerns about the reliability of their findings[92].

A further barrier to AI adoption is the lack of validation in underrepresented, resource-limited settings. While several AI systems have been evaluated in high-income countries, there is a scarcity of data from middle-income ones such as Greece and from rural or underserved regions, including parts of South Asia, sub-Saharan Africa, and Latin America. In these settings, healthcare facilities, digital systems, along with access to trained personnel or technical support may be limited, compromising AI deployment and its generalizability. Future multicenter studies should prioritize these underrepresented regions to bridge this gap and enhance global equity in AI implementation. It should be noted that most of the RCTs investigating AI applications in colonoscopy quality indicators have been conducted in East-Asian populations. This geographic imbalance may limit the generalizability of the available evidence to broader, multi-ethnic patient groups. Future large-scale, multicenter studies are needed to determine whether the pooled outcomes remain consistent across diverse clinical settings. Notably, recent Western cohort and non-Asian suggest promising cross-population applicability[85,87,93-95].

Another significant concern is cost-effectiveness. Although AI increases ADR, especially for small, diminutive adenomas and non-neoplastic polyps, it also leads to an increased number of unnecessary polypectomies, histopathologic examinations, and surveillance colonoscopies, increasing healthcare costs[20,96]. However, these costs may be limited by the long-term economic benefit of AI due to CRC prevention[20,97]. Areia et al[98] demonstrated that integrating AI in screening colonoscopy could reduce CRC incidence and mortality by 4.8% and 3.6%, respectively, in the United States population, compared to non-AI screening, resulting in yearly savings of approximately 290 million United States dollar and 57 United States dollar saved per individual. Notably, this model assumed an upfront CADe equipment cost of approximately 10000 United States dollar per unit, highlighting the favorable balance between initial investment and long-term savings. Similarly, Barkun et al[99] showed that integrating CADe in colonoscopy among FIT-positive patients in a Canadian health care setting was cost-effective, with overall savings of 14 United States dollar per patient despite the upfront cost of 1500 United States dollar per AI unit and 50 United States dollar per case. Hassan et al[100] also evaluated the cost-utility of CADe in FIT-positive populations, showing a 14.34 European dollar saved per patient compared to standard colonoscopy, with AI-assisted screening being more effective and less costly in most simulation scenarios. Collectively, these findings suggest that despite the initial costs, the integration of AI in colonoscopy could be a cost-effective or even cost-saving strategy over time, particularly when applied in organized screening programs and high-volume centers.

Nevertheless, most of these studies are based on microsimulation models, which may not fully reflect real-world intricacies and clinical challenges. Consequently, their results may not be representative of real-life scenarios. In a recent study, albeit with low-certainty evidence, it was reported that CADe adoption in screening colonoscopy provided minimal clinical benefit, increased surveillance burden and might be more suitable for endoscopists with a lower baseline performance[101]. Moreover, the goal of cost-effectiveness is most likely achieved when ADR increases from 26% to at least 30% or when the per-AI colonoscopy cost is below 579 United States dollar[102].

The integration of CADx tools can support polyp characterization, establishing novel strategies such as “diagnose and leave” or “resect and discard”, which may further reduce costs. Mori et al[103] estimated that adopting such strategies using AI could reduce mean colonoscopy costs by 149.2 million United States dollar (18.9%) in Japan, 12.3 million United States dollar (6.9%) in the United Kingdom, 1.1 million United States dollar (7.6%) in Norway, and 85.2 million United States dollar (10.9%) in the United States. Recently, it was demonstrated that combining CADe and CADx and applying these strategies in colonoscopy resulted in lower total costs (2300.76 European dollar vs 2508.75 European dollar per patient) and could avoid 173 polypectomies, 370 biopsies, and prevent 7 CRC cases per 1000 patients[104].

An important concern is the interaction between AI and humans. Overreliance on AI may result in deskilling of endoscopists, who may begin to rely more on algorithmic suggestions than on their own judgment. Although Reverberi et al[105] suggested that both expert and non-expert endoscopists are more likely to accept correct AI decisions, even when they contradict to their own initial diagnosis and reject incorrect ones, the concern remains that over time, human clinical intuition may decline. Furthermore, frequent false-positive signals from AI systems can increase procedural time and distract endoscopists, potentially limiting performance and productivity[91].

Another critical issue is patient trust and ethical responsibility. While AI provides consistency and is not compromised by fatigue and emotions, acting autonomously without human intervention raises legal and ethical concerns, especially in case of diagnostic error[106]. It remains unclear whether responsibility falls on the endoscopist, the institution, or the AI designers in case of misdiagnosis which may limit clinical adoption in the absence of well-defined legal frameworks[106]. Additionally, the use of large volumes of endoscopic video data to train and validate AI systems raises significant data privacy concerns. Compliance with privacy regulations such as the European general data protection regulation and the United States Health Insurance Portability and Accountability Act is crucial to ensure secure ethical data use, patient confidentiality and support responsible AI implementation in clinical practice. To mitigate patient anxiety and promote informed decision-making, clinicians should clearly communicate the potential for overdiagnosis associated with AI tools, especially in the context of diminutive or non-neoplastic lesions, in order to enhance transparency and support delivery care aligned with patient preferences and expectations. Moreover, AI cannot be successfully implemented without endoscopists’ acceptance and confidence. It has been reported that 93.8% of endoscopists are not convinced that AI can deliver personalized care[107]. Although most of them are optimistic about AI’s clinical utility, only 68.8% believed that AI could be integrated into routine colonoscopy and only half of them were familiar with AI technology such as neural networks and DL[107].

In low- and middle-income countries, where experienced endoscopists and advanced training programs are often limited, AI-assisted colonoscopy could serve as a cost-effective way to improve diagnostic quality. By supporting less trained staff and ensuring more consistent performance, AI tools may help reduce disparities in CRC screening and enhance care in low-resource settings.

Therefore, the successful incorporation of AI in colonoscopy will depend not only on technological progression but also on proper education, endoscopist acceptance, and well-designed, prospective studies that confirm its real-world cost-effectiveness. Adopting a synergistic human-AI interaction will be vital for increasing quality, safety, trust, and outcomes.

In terms of review methodology, as this is a narrative review, limitations include the use of a single database (PubMed) and the lack of formal quality or bias assessment, which may affect the comprehensiveness of the included literature.

CONCLUSION

AI has emerged as a breakthrough innovation in modern colonoscopy, offering significant improvements in key quality indicators, including ADR, PDR, AMR, bowel preparation assessment, WT, and cecal intubation recognition (Figure 2). By optimizing lesion characterization, supporting optical diagnosis and assisting in invasion depth prediction, AI may lead to early diagnosis and prevention of CRC. Moreover, AI integration can reshape endoscopist training by offering real-time feedback, standardizing colonoscopy quality and bridging the skill gap between novices and experts, especially in community-based or under-resourced environments. However, its widespread clinical incorporation has several challenges such us the need for high-quality, various training datasets, cost-effectiveness across healthcare systems, the risk of endoscopist deskilling and medico-legal responsibility in cases of diagnostic errors. Endoscopist acceptance and adequate education is crucial as ethical concerns and refusal to adopt new technologies may prevent its successful integration. Future research should focus on large-scale, prospective validation of AI systems under real-world conditions, assessing their performance across a wide range of quality indicators, healthcare settings, and patient populations. Additionally, combining AI with advanced imaging techniques and establishing structured, AI-driven training platforms may enhance trainee education and promote standardization of colonoscopy practice, strengthening the trust of patients and endoscopists. In parallel, clear policies addressing quality assurance, clinician education, data protection, and liability are needed to support the safe integration of AI in endoscopy. To promote adoption, healthcare systems should consider revising reimbursement models to incentivize the responsible use of AI-assisted colonoscopy. A collaborative relationship between humans and AI should be built, where AI complements rather than replaces human intelligence to ensure colonoscopy quality and provide more targeted, individualized medical care.

Figure 2 Artificial intelligence integration in key colonoscopy quality indicators. AI: Artificial intelligence; ADR: Adenoma detection rate; PDR: Polyp detection rate; APC: Adenomas per colonoscopy; AMR: Adenoma miss rate; SSLDR: Sessile serrated lesion detection rate.