Advances in artificial intelligence-based colonoscopic tools and modalities: Transforming colorectal cancer detection and management

Abstract

Colorectal cancer (CRC) remains a leading cause of cancer-related morbidity and mortality worldwide, and colonoscopy is the gold standard for screening and polyp detection. However, its effectiveness is operator-dependent, with variability in adenoma detection rates (ADR). Artificial intelligence (AI), particularly deep learning-based systems, has emerged as a promising tool to enhance colonoscopic performance. This narrative review was conducted through a structured literature search of PubMed, Scopus, and Web of Science databases for studies published up to 2025. Eligible studies included randomized controlled trials, systematic reviews, and meta-analyses evaluating AI-assisted colonoscopy, with a focus on computer-aided detection (CADe) and computer-aided diagnosis (CADx). Evidence consistently demonstrates that CADe systems improve ADR and polyp detection rate, primarily by enhancing detection of diminutive and flat lesions and reducing adenoma miss rates. CADx systems enable real-time optical diagnosis, supporting “resect-and-discard” and “diagnose-and-leave” strategies, although their diagnostic accuracy and generalizability remain variable. Beyond detection and characterization, AI applications extend to quality assurance metrics, including bowel preparation assessment and withdrawal time monitoring. Despite promising results, challenges remain, including heterogeneity among AI systems, cost-effectiveness, regulatory considerations, and integration into clinical workflows. Emerging multimodal AI approaches combining endoscopic, clinical, and molecular data may further improve risk stratification and personalized surveillance. This review provides a comprehensive synthesis of current evidence, offers critical appraisal of existing studies, and highlights future directions aimed at achieving precision-driven CRC screening and improved clinical outcomes.

Key Words: Artificial intelligence; Colonoscopy; Adenoma detection rate; Computer-aided detection; Computer-aided diagnosis; Colorectal polyps

Core Tip: Artificial intelligence (AI)-assisted colonoscopy has emerged as a promising tool for improving the quality of colorectal cancer (CRC) screening. Evidence from randomized controlled trials and meta-analyses consistently shows that computer-aided detection systems significantly increase adenoma detection rate and polyp detection rate, particularly for small and flat lesions that are commonly missed during conventional colonoscopy. However, improvements in detecting advanced adenomas and CRC remain limited. Additionally, AI may increase the removal of non-neoplastic polyps. Overall, AI technologies serve as valuable decision-support tools that enhance lesion detection, reduce operator variability, and may contribute to more effective CRC prevention strategies in clinical practice.

INTRODUCTION

Colorectal cancer (CRC) is the third most frequently diagnosed malignancy worldwide, accounting for nearly 10% of all cancer cases, and represents the second leading cause of cancer-related mortality globally[1-3]. Colonoscopy plays a central role in CRC prevention, as it reduces both incidence and mortality through the detection and removal of premalignant polyps and the early treatment of malignant lesions[4,5]. However, interval CRC may still occur, often due to low adenoma detection rates (ADRs) and incomplete polyp resection, which are recognized as major contributing factors to missed or subsequently developing cancers[1,6,7]. To address these limitations, several strategies have been introduced to improve ADR, including educational programs for endoscopists, the implementation of advanced imaging technologies, and the use of mechanical devices designed to enhance mucosal visualization. In recent years, artificial intelligence (AI)-based computer-aided diagnosis (CADx) systems have emerged as a promising technological advancement with the potential to further enhance adenoma detection during colonoscopy[1,8]. These systems have been specifically developed to assist endoscopists in both the detection of polyps through computer-aided detection (CADe) and the characterization of lesions using CADx during colonoscopic procedures[8-10]. Overall, AI-driven technologies represent an important advancement in endoscopic practice, showing considerable potential to improve key quality indicators, enhance procedural standardization, and reduce variability between operators.

The aim of this review is to evaluate the current evidence regarding the role of AI-assisted colonoscopy in improving colorectal polyp detection and colonoscopy quality indicators. Specifically, this paper summarizes findings from randomized controlled trials and systematic reviews assessing the impact of CADe and diagnosis systems on ADRs, polyp detection rate (PDR), and other clinically relevant outcomes. Additionally, the review discusses the potential benefits, limitations, and future directions of AI integration in routine colonoscopic practice.

Evidence from systematic reviews and meta-analyses

AI applications in colonoscopy are broadly categorized into CADe and CADx, which serve distinct but complementary roles. CADe systems are designed to assist in real-time identification of colorectal lesions by highlighting suspicious mucosal abnormalities during endoscopy. These systems primarily aim to reduce perceptual errors and improve ADR, particularly for small, flat, or subtle lesions. In contrast, CADx systems focus on lesion characterization, providing real-time histologic prediction based on endoscopic imaging. CADx supports strategies such as “resect-and-discard” for diminutive adenomas and “diagnose-and-leave” for non-neoplastic lesions, potentially reducing unnecessary polypectomies and pathology costs. While CADe has demonstrated consistent improvements in detection outcomes, CADx performance remains more variable, with limitations in diagnostic accuracy and generalizability. Integration of both systems into a unified workflow represents a key future direction in AI-assisted colonoscopy.

Several systematic reviews and meta-analyses have consistently demonstrated that AI-assisted colonoscopy improves key colonoscopy quality indicators, particularly ADR and PDR. However, the magnitude of benefit and its impact on clinically significant lesions remain areas of ongoing investigation.

Impact of AI on ADR

Rather than reiterating individual study findings in detail, Table 1 summarizes the key outcomes of major meta-analyses evaluating AI-assisted colonoscopy. Collectively, these studies demonstrate a consistent improvement in ADR and PDR, with relative increases ranging from 20% to 50%. The benefit is primarily driven by increased detection of diminutive and nonadvanced lesions, while improvements in advanced neoplasia detection remain limited.

Table 1 Summary of systematic reviews and meta-analyses evaluating the impact of artificial intelligence-assisted colonoscopy on adenoma detection rate.

Ref.	Study type	Number of studies/participants	Main outcomes	Key findings
Khalaf et al[10], 2025	Systematic review and meta-analysis	7 RCTs/5427 participants	ADR, PDR	AI-assisted colonoscopy significantly increased ADR and PDR. Most additional lesions detected were small polyps, while detection of advanced adenomas and pedunculated lesions was lower
Huang et al[15], 2022	Meta-analysis	10 RCTs/6629 participants	ADR, PDR, lesions per colonoscopy	AI significantly improved ADR and PDR. Higher numbers of adenomas, polyps, and sessile serrated lesions were detected per colonoscopy, with benefits across most lesion sizes and locations except pedunculated lesions and cecal lesions
Makar et al[11], 2025	Updated meta-analysis	28 RCTs/23861 participants	ADR, adenoma miss rate	CADe increased ADR by approximately 20% and reduced adenoma miss rates by 55%. Benefits were consistent across platforms, endoscopist experience levels, and clinical settings
Hassan et al[16], 2021	Meta-analysis	5 RCTs/4354 participants	ADR	CADe consistently demonstrated benefit (36.6% vs 25.2%) compared with conventional colonoscopy, with improvements across different lesion sizes, morphologies, and anatomical locations
Hassan et al[12], 2023	Systematic review and meta-analysis	21 RCTs/18232 participants	ADR, adenoma miss rate	CADe-assisted colonoscopy significantly increased ADR (44.0% vs 35.9%) and reduced adenoma miss rates compared with standard colonoscopy
Soleymanjahi et al[13], 2024	Systematic review and meta-analysis	44 RCTs/36201 participants	ADR, adenomas per colonoscopy	CADe was associated with increased detection and increased the average number of adenomas detected per colonoscopy
Adiwinata et al[17], 2023	Systematic review and meta-analysis	Multiple studies	ADR	AI-assisted colonoscopy significantly improved ADR with a pooled odds ratio of 1.58, confirming a clear detection advantage over conventional colonoscopy

AI: Artificial intelligence; ADR: Adenoma detection rate; PDR: Polyp detection rate; RCTs: Randomised controlled trials; CADe: Computer-aided detection.

Overall, the cumulative evidence from randomized controlled trials and meta-analyses consistently demonstrates that AI-assisted colonoscopy significantly improves ADR, with relative increases ranging from approximately 20% to 50% compared with conventional colonoscopy. Importantly, these improvements are predominantly driven by enhanced detection of diminutive (≤ 5 mm) and nonadvanced adenomas, while detection rates of advanced neoplasia remain largely unchanged. This suggests that AI primarily functions as a detection-enhancing tool rather than a risk-stratification modality. Furthermore, reductions in adenoma miss rates—reported to be as high as 50%-55% highlight the role of AI in minimizing perceptual errors during mucosal inspection. Collectively, these findings support the integration of CADe as an adjunct to improve procedural quality; however, the clinical significance of increased detection of diminutive lesions remains an area of ongoing debate[11-17].

Although the improvement in ADR is one of the most consistent findings across studies, we believe that the clinical significance of this increase should be interpreted with caution. The predominance of diminutive adenomas among newly detected lesions suggests that AI may be more effective at enhancing visual sensitivity rather than altering clinically meaningful outcomes. Future research should therefore focus on whether these improvements translate into reductions in interval CRC and mortality.

Impact of AI on PDR

In addition to improving ADR, AI-assisted colonoscopy has consistently demonstrated improvements in PDR.

A meta-analysis of five randomized prospective trials reported that AI-assisted colonoscopy significantly improved both ADR and PDR with high certainty of evidence. The benefit was primarily driven by increased detection of small (≤ 5 mm) nonadvanced adenomas, while detection of larger or advanced adenomas remained unchanged[18]. Another meta-analysis including 12 randomized trials with 11267 participants demonstrated that AI significantly improved PDR, although the improvement in ADR did not reach statistical significance due to heterogeneity among studies[19].

A systematic review evaluating 17 randomized controlled trials published between 2015 and 2025 also confirmed consistent improvements in both ADR and PDR with AI-assisted colonoscopy. Detection accuracy exceeded 85% for adenomas and 90% for polyps, with the greatest benefit observed in identifying small and flat lesions that are frequently overlooked during conventional procedures[20].

The observed increase in PDR further supports the role of AI in reducing perceptual errors during colonoscopy. However, the concomitant rise in detection of non-neoplastic lesions highlights a potential trade-off between sensitivity and specificity, which may have implications for procedural efficiency and healthcare costs.

Detection of advanced neoplasia and sessile serrated lesions

Despite clear improvements in overall lesion detection, several studies suggest that AI-assisted colonoscopy provides limited benefit in detecting advanced neoplasia.

Many meta-analyses have shown that the increase in detection is largely driven by diminutive and nonadvanced adenomas rather than clinically significant lesions. For example, the previously mentioned meta-analysis of 28 randomized trials reported no significant improvement in advanced adenoma detection despite increases in ADR[11]. Similarly, Soleymanjahi et al[13], found that detection of advanced colorectal neoplasia remained comparable between AI-assisted and conventional colonoscopy, although the overall number of adenomas detected increased.

However, some studies suggest modest improvements in detecting specific lesion types. For instance, Huang et al[15], meta-analysis reported higher detection of sessile serrated lesions with AI-assisted colonoscopy.

The limited impact of AI on advanced neoplasia detection underscores a critical limitation of current systems. In our opinion, this reflects the need for next-generation algorithms capable not only of detecting lesions but also of prioritizing clinically significant pathology.

Comparative performance of AI systems

Differences in performance among AI platforms have also been investigated (Table 2). A large systematic review including 64 studies and more than 50000 patients demonstrated that AI-assisted colonoscopy was associated with increased detection compared with conventional colonoscopy. Among available systems, endoangel showed the highest effectiveness for detecting adenomas and polyps, while endocuff-AI performed best in detecting sessile serrated lesions. However, CADx systems did not significantly improve sensitivity or specificity for distinguishing neoplastic polyps[21]. Another meta-analysis evaluating different AI platforms also reported variability in system performance. Endoangel demonstrated the highest efficacy, followed by EndoAID, while CADx eye and gastrointestinal (GI) genius showed comparable detection outcomes[22].

Table 2 Summary of randomized controlled trials assessing the clinical performance of artificial intelligence-assisted colonoscopy.

Ref.	Study design	Participants	AI system/intervention	Main outcomes	Key findings
Repici et al[23], 2020	Multicenter randomized trial	685 patients	GI-genius CADe system	ADR, adenomas per colonoscopy	AI-assisted colonoscopy significantly increased ADR (54.8% vs 40.4%) and adenomas detected per colonoscopy, mainly due to improved detection of small adenomas ≤ 9 mm without increasing withdrawal time
Seager et al[24], 2024	Multicenter randomized control trial (COLO-DETECT trial)	Not specified in text	GI-genius CADe system	ADR, adenomas per procedure	AI significantly improved ADR (56.6% vs 48.4%) and mean adenomas per procedure without increasing adverse events
Nakashima et al[25], 2023	Randomized trial	415 patients	CADe-assisted colonoscopy	ADR, adenoma miss rate	Computer-aided detection improved ADR (59.4% vs 47.6%) and reduced adenoma miss rates in the rectosigmoid colon without prolonging examination time
Wang et al[26], 2020	Double-blind randomized trial	962 patients	CADe system	ADR	ADR significantly improved (34% vs 28%). AI particularly enhanced detection of subtle lesions such as small, flat, or partially hidden polyps
Lau et al[27], 2024	Randomized trial	766 patients	Computer-aided detection -assisted colonoscopy	ADR	CADe significantly increased ADR among endoscopists-in-training and improved detection of small adenomas in both right and left colon
Xu et al[28], 2021	Randomized trial	2352 patients	AI-assisted colonoscopy	Polyps per colonoscopy	AI increased detection of additional polyps per colonoscopy and improved identification of diminutive and flat lesions
Glissen Brown et al[29], 2022	Tandem colonoscopy randomized contral trial	Not specified in text	CADe-system	Adenoma miss rate	CADe significantly reduced adenoma miss rate (20.1% vs 31.3%) and increased adenomas detected per colonoscopy
Luo et al[30], 2021	Randomized back-to-back colonoscopy study	150 patients	Real-time AI polyp detection system	PDR	AI significantly improved polyp detection rate, mainly through improved detection of diminutive polyps

AI: Artificial intelligence; ADR: Adenoma detection rate; PDR: Polyp detection rate; CADe: Computer-aided detection; GI: Gastrointestinal.

The variability observed among AI platforms suggests that system performance is highly dependent on algorithm design and training data. This heterogeneity limits direct comparability and highlights the need for standardized benchmarking frameworks.

Procedural effects and detection trade-offs

Although AI-assisted colonoscopy improves lesion detection, several studies have reported procedural trade-offs.

For example, Makar et al[11], reported slightly longer withdrawal times when CADe was used. Additionally, the technology increased the removal of non-neoplastic polyps, raising concerns about potential overdiagnosis and unnecessary resections. Similarly, Hassan et al[12] found that CADe increased the number of non-neoplastic polyps removed without improving detection of advanced adenomas.

Overall, while AI-assisted colonoscopy significantly enhances detection of colorectal lesions, the clinical implications of detecting large numbers of diminutive polyps remain an important consideration.

CRITICAL APPRAISAL OF CURRENT EVIDENCE

Despite robust evidence supporting AI-assisted colonoscopy, several methodological limitations must be acknowledged. First, significant heterogeneity exists across studies, including variability in patient populations, endoscopist expertise, bowel preparation quality, and withdrawal times—all of which independently influence ADR. Second, most randomized trials were conducted in high-volume centers or expert settings, potentially limiting generalizability to community practice. Third, variability among AI systems including differences in training datasets, algorithm architectures, and real-time processing capabilities—makes direct comparison between platforms challenging. Network meta-analyses suggest differential performance among systems, but these findings should be interpreted cautiously due to indirect comparisons and inconsistent outcome definitions[21,22].

Additionally, performance bias may arise from the “Hawthorne effect”, whereby endoscopists modify behavior when assisted by AI, potentially inflating observed benefits. Detection bias is also relevant, as most studies are not blinded to intervention allocation.

Finally, outcome measures such as ADR, while validated, may not fully capture clinically meaningful endpoints such as interval CRC reduction. Long-term outcome data remain limited, highlighting the need for prospective studies evaluating cancer incidence and mortality.

Evidence from randomized controlled trials

Multiple randomized controlled trials have evaluated the real-world performance of AI-assisted colonoscopy.

A multicenter randomized trial involving 685 patients demonstrated that a deep learning-based CADe system (GI-genius) led to increased ADR detection (54.8% vs 40.4%) and increased the number of adenomas detected per colonoscopy. The improvement was mainly attributed to enhanced detection of small adenomas ≤ 9 mm without increasing withdrawal time[23].

Similarly, the COLO-DETECT trial conducted across 12 hospitals in England showed that the GI genius system significantly improved detection outcomes, increasing both ADR (56.6% vs 48.4%) and mean adenomas per procedure without increasing adverse events[24]. Another randomized study including 415 participants found that CADe-assisted colonoscopy improved ADR detection (59.4% vs 47.6%) and reduced adenoma miss rates in the rectosigmoid colon without prolonging examination time[25].

A double-blind randomized trial involving 962 patients further confirmed that CADe is beneficial in ADR detection (34% vs 28%), particularly by detecting subtle lesions such as small, flat, or partially hidden polyps[26]. Additional studies have demonstrated similar benefits. A randomized study involving 766 patients showed that CADe used among endoscopists-in-training was associated with increased detection of small adenomas and lesions in both the right and left colon[27].

Another randomized trial involving 2352 patients demonstrated that AI assistance increased detection of additional polyps per colonoscopy and improved identification of diminutive and flat lesions[28]. A tandem colonoscopy trial also demonstrated that CADe significantly reduced adenoma miss rates (20.1% vs 31.3%) and increased adenomas detected per colonoscopy[29]. Similarly, a randomized back-to-back colonoscopy study involving 150 patients showed that AI significantly improved PDRs, primarily due to better identification of diminutive polyps[30].

Randomized controlled trials provide strong evidence supporting AI-assisted colonoscopy; however, most studies are conducted in controlled environments. We believe that real-world validation in diverse clinical settings is essential before widespread implementation can be fully justified.

Studies showing limited or variable benefit

Despite promising results, some studies have reported more modest or inconsistent improvements.

The AI-SEE randomized trial conducted in four United States community endoscopy centers found no significant improvement in ADR or adenomas per colonoscopy with CADe use, although detection of nonadenomatous polyps increased and withdrawal time was slightly longer[31]. Similarly, another prospective randomized tandem colonoscopy study conducted in three Asian centers reported that CADe did not significantly reduce proximal adenoma miss rates, although detection during the first examination improved[32].

A randomized trial conducted in a Spanish CRC screening program also reported that CADe did not significantly improve detection of advanced colorectal neoplasia or overall ADR, although detection of small and proximal lesions increased[33]. Alali et al[9], showed improved polyp detection with CADe but no statistically significant difference in ADR, likely due to limited sample size.

Finally, a pilot study comparing CADx-assisted colonoscopy with historical controls showed only a non-significant increase in adenomas per colonoscopy, although detection of non-neoplastic polyps significantly increased[34].

The presence of studies demonstrating limited or no benefit emphasizes that AI is not universally effective. These findings suggest that baseline endoscopist performance and procedural quality may influence the magnitude of benefit derived from AI assistance.

Role of CADx in polyp characterization

In contrast to CADe systems, which focus on lesion detection, CADx systems aim to characterize detected polyps by predicting histology in real time. This capability enables the implementation of “resect-and-discard” strategies for diminutive adenomas and “diagnose-and-leave” approaches for non-neoplastic lesions, potentially reducing unnecessary polypectomies and pathology costs[1,16,35].

Several studies have evaluated the diagnostic performance of CADx systems using advanced imaging modalities such as narrow-band imaging. A large multicenter study by Rex et al[35] demonstrated that CADx achieved high specificity for adenoma prediction but did not significantly improve sensitivity compared with expert endoscopists. These findings suggest that CADx may be useful as a decision-support tool but is not yet reliable for independent clinical decision-making. Importantly, the Preservation and Incorporation of valuable endoscopic innovations (PIVI) thresholds established by the American Society for Gastrointestinal Endoscopy require ≥ 90% agreement in surveillance interval assignment for adoption of “resect-and-discard” strategies. Current CADx systems approach but do not consistently meet these thresholds across all clinical settings[35]. Challenges associated with CADx include variability in imaging quality, differences in lesion morphology, and limited generalizability across populations. Additionally, most systems are trained on high-quality datasets, which may not reflect real-world conditions.

Despite these limitations, CADx represents a critical step toward precision endoscopy. Future developments integrating CADx with CADe systems and multimodal data sources may enable fully automated detection-diagnosis pipelines, improving both efficiency and clinical decision-making. While CADx represents a promising step toward real-time optical diagnosis, current evidence suggests that its performance remains insufficient for independent clinical decision-making. Further refinement and validation are required before widespread adoption of “resect-and-discard” strategies can be recommended.

LIMITATIONS AND FUTURE DIRECTIONS

Despite encouraging results, several limitations should be considered. First, most studies demonstrate improvements primarily in the detection of diminutive or nonadvanced adenomas rather than advanced colorectal neoplasia. The clinical significance of detecting large numbers of small polyps remains uncertain.

Second, AI-assisted colonoscopy may increase the detection and removal of non-neoplastic polyps, potentially leading to unnecessary resections and increased healthcare costs. Additionally, small increases in withdrawal time have been reported in some studies.

Third, heterogeneity among AI systems, study populations, endoscopist expertise, and colonoscopy techniques makes direct comparisons difficult. Differences in training datasets, algorithm design, and hardware platforms may influence system performance.

Finally, long-term clinical outcomes such as reductions in CRC incidence and mortality have not yet been conclusively demonstrated. Future research should focus on large multicenter trials, real-world implementation studies, and improvements in CADx systems to enhance histologic characterization.

Beyond technical limitations, several real-world challenges must be considered. Cost-effectiveness remains a critical issue, as AI systems require substantial investment in hardware, software integration, and maintenance. While some modeling studies suggest that increased ADR may offset costs through reduced CRC incidence, real-world economic data are still limited. Implementation barriers include workflow integration, training requirements, and potential alert fatigue due to false-positive detections. Additionally, regulatory approval processes vary across regions, with agencies such as the United States Food and Drug Administration and European CE marking requiring rigorous validation of safety and effectiveness. Ethical and legal considerations are also emerging, including data privacy, algorithm transparency, and medico-legal liability in cases of missed lesions despite AI assistance. These factors must be addressed to facilitate widespread adoption in routine clinical practice[2,16,36].

Recent guidelines and expert consensus statements have begun to address the role of AI in colonoscopy. The European Society of Gastrointestinal Endoscopy acknowledges that CADe systems improve ADR and may be considered as an adjunct to standard colonoscopy, although routine implementation is not yet universally recommended due to limited long-term outcome data[12]. Similarly, the American Society for Gastrointestinal Endoscopy has recognized AI as a promising tool for enhancing quality metrics, particularly in polyp detection and optical diagnosis, while emphasizing the need for standardized validation and real-world studies before widespread adoption[37]. These evolving recommendations highlight the transitional phase of AI integration into endoscopic practice.

CONCLUSION

AI-assisted colonoscopy represents a major advancement in CRC screening, with strong evidence supporting improvements in adenoma and PDRs. However, the clinical impact of these improvements on advanced neoplasia detection and long-term outcomes remains uncertain (Figure 1). Future research should focus on large multicenter trials and real-world implementation studies to validate the effectiveness of AI across diverse clinical settings. In particular, the development of advanced CADx systems capable of reliable histologic prediction remains a key priority. Emerging multimodal AI approaches integrating endoscopic imaging with clinical, genetic, and molecular data offer a promising pathway toward personalized CRC prevention. These systems may enable more accurate risk stratification and tailored surveillance strategies. Ultimately, long-term studies evaluating reductions in CRC incidence and mortality will be essential to establish the true clinical value of AI-assisted colonoscopy and to guide its integration into routine practice.

Figure 1 Clinical impact of artificial intelligence-assisted colonoscopy. AI: Artificial intelligence; ADR: Adenoma detection rate; PDR: Polyp detection rate.