Artificial intelligence for reducing missed detection of adenomas and polyps in colonoscopy: A systematic review and meta-analysis

Abstract

BACKGROUND

Colorectal cancer has a high incidence and mortality rate, and the effectiveness of routine colonoscopy largely depends on the endoscopist’s expertise. In recent years, computer-aided detection (CADe) systems have been increasingly integrated into colonoscopy to improve detection accuracy. However, while most studies have focused on adenoma detection rate (ADR) as the primary outcome, the more sensitive adenoma miss rate (AMR) has been less frequently analyzed.

AIM

To evaluate the effectiveness of CADe in colonoscopy and assess the advantages of AMR over ADR.

METHODS

A comprehensive literature search was conducted in PubMed, Embase, and the Cochrane Central Register of Controlled Trials using predefined search strategies to identify relevant studies published up to August 2, 2024. Statistical analyses were performed to compare outcomes between groups, and potential publication bias was assessed using funnel plots. The quality of the included studies was evaluated using the Cochrane Risk of Bias tool and the Grading of Recommendations, Assessment, Development, and Evaluation approach.

RESULTS

Five studies comprising 1624 patients met the inclusion criteria. AMR was significantly lower in the CADe-assisted group than in the routine colonoscopy group (147/927, 15.9% vs 345/960, 35.9%; P < 0.01). However, CADe did not provide a significant advantage in detecting advanced adenomas or lesions measuring 6-9 mm or ≥ 10 mm. The polyp miss rate (PMR) was also lower in the CADe-assisted group [odds ratio (OR), 0.35; 95% confidence interval (CI): 0.23-0.52; P < 0.01]. While the overall ADR did not differ significantly between groups, the ADR during the first-pass examination was higher in the CADe-assisted group (OR, 1.37; 95%CI: 1.10-1.69; P = 0.004). The level of evidence for the included randomized controlled trials was graded as moderate.

CONCLUSION

CADe can significantly reduce AMR and PMR while improving ADR during initial detection, demonstrating its potential to enhance colonoscopy performance. These findings highlight the value of CADe in improving the detection of colorectal neoplasms, particularly small and histologically distinct adenomas.

Key Words: Artificial intelligence; Computer-aided detection; Colonoscopy; Neoplasms; Prevention and control

Core Tip: Artificial intelligence is being increasingly used in colonoscopy, with more and more studies reporting its potential benefits. However, most studies have focused on adenoma detection rate (ADR) as the primary outcome and assessed only short-term effects. Recently, adenoma miss rate (AMR) has gained more attention, and based on this, we designed this meta-analysis to evaluate the effect of computer-aided detection on AMR, compared it with ADR, and assessed its long-term impact.

INTRODUCTION

Colorectal cancer (CRC) is the fourth most common malignancy worldwide, accounting for 6.1% of all cancer cases, and is the second leading cause of cancer-related mortality (9.2%) after lung cancer[1]. Although colonoscopy remains one of the most effective methods for diagnosing and managing gastrointestinal diseases[2], its effectiveness is largely dependent on the endoscopist’s skill, and the low detection rate of adenomas and polyps in conventional colonoscopy contributes to the high incidence of CRC, increasing both patient burden and healthcare costs[3].

With advancements in medical technology, artificial intelligence (AI) has been increasingly integrated into clinical practice. AI-assisted colonoscopy has demonstrated high diagnostic accuracy, with one study reporting an accuracy of 98% and a specificity of 100% for detecting colorectal neoplasms[4]. By enhancing visual recognition, AI can improve the detection of both subtle and advanced adenomas, thereby reducing the risk of misclassifying non-neoplastic lesions. Thus, it may help lower the rate of unnecessary resections and decrease complications such as intestinal bleeding and perforation[5,6].

In 2019, the computer-aided polyp detection (CADe) system was introduced into clinical practice, and it demonstrated a significant improvement in adenoma detection rate (ADR) compared to conventional colonoscopy (29.1% vs 20.3%, P < 0.001)[7]. This system is based on convolutional neural networks (CNNs) or self-learning algorithms that integrate visual recognition with diagnostic criteria, providing real-time alerts to endoscopists regarding suspicious lesions.

Until now, most studies that have evaluated AI-assisted colonoscopy have used ADR as the primary outcome measure[8,9]. However, AI-based systems may generate false positives by detecting non-adenomatous lesions, leading to an overestimation of ADR[10,11]. To address this limitation, adenoma miss rate (AMR) has been proposed as a more reliable metric, as it accounts for lesions that are missed during the initial examination and are subsequently detected upon repeat colonoscopy. AMR provides a more accurate assessment of AI performance, as it directly reflects the system’s ability to reduce missed lesions. While previous studies have explored AMR in colonoscopy, most have been limited to short-term evaluations due to time constraints[12-14].

Herein, we designed this study using latest research to determine the long-term effects of AI in colonoscopy, which have not been extensively explored in previous studies. Specifically, we analyzed the impact of the CADe system on AMR and provided a comprehensive overview of advancements in AI-assisted colorectal disease detection. By highlighting the advantages of AMR over ADR, this study further demonstrates the potential of AI to enhance colonoscopy performance and improve diagnostic accuracy.

MATERIALS AND METHODS

Data sources and search strategy

A comprehensive literature search was conducted in PubMed, Embase, and the Cochrane Central Register of Controlled Trials to identify studies published up to August 2, 2024. The search was restricted to randomized controlled trials (RCTs), although studies involving the CADe system were also reviewed. The detailed search strategy is provided in Supplementary materials, and the study flow chart is illustrated in Figure 1.

Figure 1 Study flow chart. CENTRAL: Cochrane Central Register of Controlled Trials; CADe: Computer-aided detection.

Outcomes and inclusion and exclusion criteria

Primary outcome: The primary outcome was the AMR, defined as the number of adenomas detected during the second-pass colonoscopy divided by the total number of adenomas identified across both passes. Subgroup analyses were conducted based on adenoma size, location, and histological characteristics.

Secondary outcomes: The secondary outcomes included the polyp miss rate (PMR), ADR, adenomas per colonoscopy (APC), and withdrawal time. The PMR was calculated as the number of polyps detected during the second-pass colonoscopy divided by the total number of polyps detected across both passes. ADR was defined as the proportion of individuals undergoing a complete colonoscopy in whom at least one adenoma was detected, APC was determined by dividing the total number of adenomas detected by the total number of colonoscopies performed, and withdrawal time was measured as the duration required to withdraw the colonoscope while examining the colonic mucosa, excluding the time needed for biopsy or lesion excision.

Inclusion and exclusion criteria: Studies were included if they enrolled adults aged 18 years or older who underwent colonoscopy, while those that were not RCTs, did not report relevant outcomes, or had a sample size of fewer than 30 participants were excluded.

Selection process

Two independent reviewers (Wang SY and Gao JC) screened the search results based on predefined criteria. The screening process involved an initial review of titles and abstracts, followed by a full-text assessment of studies meeting the inclusion criteria. All screening decisions were documented, and any discrepancies were resolved through discussion, and there were no unresolved disagreements by the time of manuscript completion.

Data extraction

Data were independently extracted by Wang SY and Gao JC using standardized tables comprising study characteristics (author, country, year, and study design), patient characteristics (sample size, age, and sex), and outcome measures (AMR, PMR, ADR, APC, and withdrawal time). In cases where data were missing, the studies’ corresponding authors were contacted, but no responses were received. Any discrepancies in data extraction were resolved through discussion with a third reviewer (Wu SD).

Study quality and assessment

The quality of the included studies was assessed using the Cochrane Risk of Bias tool, and the results are summarized in Figure 2.

Figure 2 Study quality and assessment chart.

Data synthesis and statistical analysis

Dichotomous outcomes were analyzed using odds ratios (OR) with 95% confidence interval (CI), while mean differences (MD) with 95%CI were used for continuous outcomes. The DerSimonian and Laird random-effects model was applied for all analyses. Given the skewed distribution of continuous variables in the included studies, this study applied the methods proposed by Luo et al[15] and Wan et al[16] to estimate mean values and standard deviations. Subgroup analyses were conducted to evaluate missed adenomas based on size (≤ 5 mm, 6-9 mm, and ≥ 10 mm), location (proximal colon and distal colon), and histological characteristics (sessile serrated lesions and advanced adenomas). To assess the sensitivity of ADR in evaluating CADe performance, subgroup analyses were performed for all detected adenomas, including ADR and first-pass ADR. Withdrawal time was also analyzed in subgroups based on first-pass and second-pass examinations.

Heterogeneity was assessed using Cochran’s Q (χ²) and I² statistics, whereby I² values of 25%, 50% and 75% indicated low, moderate, and high heterogeneity, respectively. Sensitivity analysis for the primary outcome (AMR) was performed using the leave-one-out method. Publication bias was evaluated using funnel plots.

The quality of evidence was graded using the Grading of Recommendations, Assessment, Development, and Evaluation methodology[17]. This meta-analysis adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines[18]. All statistical analyses were conducted using RevMan 5.4 (The Cochrane Collaboration, Copenhagen, Denmark). The meta-analysis was registered in PROSPERO (ID: CRD42024583571).

RESULTS

A total of 738 records were identified through the database search. After screening, five RCTs met the inclusion criteria, comprising a total of 1624 patients[19-23]. All five studies were tandem colonoscopy trials in which the CADe-assisted groups utilized the CADe system. However, one study incorporated an additional real-time assistance system, the CAQ system, which may have contributed to an increased false positive rate. Of these studies, three were conducted in Asia[19,20,23], one in North America[21], and one in Europe[22]. The key characteristics of these studies are summarized in Table 1.

Table 1 Study characteristics, n (%).

Ref.	Year	Sample size	Age (mean years)		Gender (male)		Country	AI system	Endoscopists	Indication			Bowel preparation scale
			CADe	Routine	CADe	Routine				CADe		Routine	CADe	Routine
Wang et al[19]	2020	369	47.72 ± 10.82	47.19 ± 10.38	93 (50.54)	86 (46.69)	China	CADe (EndoScreener)	Three experienced endoscopists	Screening: 58; ymptomatic: 107; Surveillance: 19	Screening: 55; Symptomatic: 117; Surveillance: 13		7.11 ± 1.40	7.19 ± 1.42
Kamba et al[10]	2021	355	61.63 ± 9.89	61.44 ± 10.01	136 (76.40)	136 (76.80)	Japan	CADe (Locally system)	Experts (> 5000 colonscopies) and non-experts (< 5000)	Screening: 88; Surveillance: 59; Other (FOBT positive): 31	Screening: 78; Surveillance: 68; Other (FOBT positive): 31		1-3: 166; 4-5: 5	1-3: 169; 4-5: 4
Glissen Brown et al[21]	2022	223	61.18 ± 9.83	60.51 ± 8.45	54 (47.79)	68 (61.82)	United States	CADe (EndoScreener)	_	Screening: 68; Surveillance: 45	Screening: 65; Surveillance: 45		9.00 (8.00-9.00)	9.00 (8.00-9.00)
Wallace et al[22]	2022	230	63.00 ± 8.20	64.60 ± 8.10	80 (68.97)	77 (67.54)	Italy, United Kingdom, and United States	CADe (GI-Genius)	Endoscopists (> 1000 colonscopies)	Screening: 41; Surveillance: 75	Screening: 41; Surveillance: 73		8.03 ± 1.27	8.08 ± 1.46
Yao et al[23]	2024	456	50.63 ± 12.25	49.85 ± 11.68	117 (51.54)	123 (53.70)	China	CADe + CAQ (EndoAngel)	8 novice endoscopists (< 5000 colonscopies) and 10 expert endoscopists (> 5000)	Screening: 145; Symptomatic: 60; Surveillance: 22	Screening: 146; Symptomatic: 64; Surveillance: 19		Inadequate (< 6): 31; Adequate (> 6): 196	Inadequate (< 6): 27; Adequate (> 6): 202

CADe: Computer-aided detection; CAQ: Computer-aided quality improvement; FOBT: Fecal occult blood test.

Among the included studies, only one implemented a double-blind design, which was assessed as having a low risk of performance and detection bias[23]. The remaining studies used a single-blind design. In one study, the randomization method and the implementation of allocation concealment were not explicitly stated[22]. The risk of bias assessment for all included studies is presented in Figure 2.

AMR

A total of five studies were included in the analysis, all of which concluded that the CADe system significantly reduced AMR. We also observed that AMR was lower in the CADe-assisted group compared to the routine colonoscopy group (147/927, 15.9% vs 345/960, 35.9%; OR, 0.34; 95%CI: 0.26-0.45; P < 0.01). Heterogeneity was assessed as moderate (I² = 35%) (Figure 3A). No significant publication bias was detected in the funnel plot (Figure 3B). Sensitivity analysis demonstrated that the exclusion of the study by Glissen Brown et al[21] reduced heterogeneity to 0 (Figure 3C).

Figure 3 Forrest plots and funnel plot. A: Forrest plot showing adenoma miss rate for colonoscopy with vs without computer-aided detection (CADe) assistance for the included studies; B: Funnel plot showing associated publication bias; C: Sensitivity analysis of adenoma miss rate for colonoscopy with vs without CADe assistance for the included studies. CADe: Computer-aided detection; OR: Odds ratio.

In the morphological subgroup analysis, all five studies provided data comparing AMR between the CADe-assisted and routine colonoscopy groups for sessile serrated lesions. The AMR was significantly lower in the CADe-assisted group (4/48, 8.3% vs 26/65, 40.0%; OR, 0.16; 95%CI: 0.05-0.50; P = 0.001), with heterogeneity assessed as low (I² = 0) (Figure 4A). However, for advanced adenomas, the difference between the CADe-assisted and routine groups was not statistically significant (7/47, 14.9% vs 27/58, 46.6%; OR, 0.35; 95%CI: 0.07-1.69; P = 0.19), and heterogeneity was moderate (I² = 48%) (Figure 4B). Subgroup analysis based on adenoma size and location using data from two studies showed that in the ≤ 5 mm subgroup, the AMR was significantly lower in the CADe-assisted group compared to the routine colonoscopy group (OR, 0.33; 95%CI: 0.22-0.50; P < 0.01). However, no significant differences were observed in the 6-9 mm (OR, 0.67; 95%CI: 0.28-1.62; P = 0.37) or ≥ 10 mm subgroups (OR, 0.21; 95%CI: 0.04-1.07; P = 0.06). Heterogeneity in this analysis was low (I² = 0) (Figure 4C).

Figure 4 Forrest plots showing adenoma miss rate for colonoscopy with vs without computer-aided detection assistance. A: Sessile serrated lesions for the included studies; B: Advanced adenomas for the included studies; C: Size for the included studies; D: Location for the included studies. CADe: Computer-aided detection.

Regarding adenoma location, the AMR was significantly lower in the CADe-assisted group than in the routine group in both the proximal colon (OR, 0.48; 95%CI: 0.30-0.76; P = 0.002) and distal colon (OR, 0.20; 95%CI: 0.11-0.39; P < 0.01). Heterogeneity was low (I² = 0) (Figure 4D).

PMR, ADR, and APC

A total of five studies assessed the PMR, and all of them reported a significantly lower PMR in the CADe-assisted group compared to the routine colonoscopy group (330/1, 848, 17.9% vs 693/1, 863, 37.2%; OR, 0.35; 95%CI: 0.23-0.52; P < 0.01), and the heterogeneity was assessed as high (I² = 85%) (Figure 5). Since sensitivity analysis revealed that the exclusion of any single study did not substantially reduce heterogeneity, no subgroup analysis was performed.

Figure 5 Forrest plot showing polyps miss rate for colonoscopy with vs without computer-aided detection assistance for the included studies. CADe: Computer-aided detection.

For ADR, data from four studies were analyzed as one study did not provide relevant data. The results showed no significant difference between the CADe-assisted and routine colonoscopy groups (42.3% vs 41.3%; OR, 1.03; 95%CI: 0.79-1.34; P = 0.83; I² = 18%) (Figure 6A). However, when all five studies were included, and first-pass ADR was specifically analyzed, the CADe-assisted group showed a significant improvement (43.2% vs 36.5%; OR, 1.37; 95%CI: 1.10-1.69; P = 0.004; I² = 0) (Figure 6B). Both analyses demonstrated low heterogeneity.

Figure 6 Forrest plots showing adenoma detection rate for colonoscopy with vs without computer-aided detection assistance. A: The included studies; B: The first pass for the included studies. CADe: Computer-aided detection.

For APC, data from four studies were analyzed, and no significant difference was found between the CADe-assisted and routine colonoscopy groups (MD, 0.07; 95%CI: -0.12 to 0.25; P = 0.47; I² = 24%) (Figure 7). The heterogeneity in this analysis was low.

Figure 7 Forrest plot showing adenomas per colonoscopy for colonoscopy with vs without computer-aided detection assistance for the included studies. CADe: Computer-aided detection.

Withdrawal time(s)

Since the calculation of the missed detection rate requires two separate colonoscopies, withdrawal times for both procedures were analyzed separately. The withdrawal time was measured in seconds, and data from four studies were included. No significant difference was observed between the CADe-assisted and routine colonoscopy groups for either the first procedure (MD, 33.53; 95%CI: -19.14 to 86.20; P = 0.21; I² = 94%) or the second procedure (MD, 6.33; 95%CI: -24.69 to 37.35; P = 0.69; I² = 89%) (Figure 8A and Figure 9). Both analyses demonstrated high heterogeneity. For the first procedure, heterogeneity was substantially reduced following the exclusion of the study by Glissen Brown et al[21] (Figure 8B). However, for the second procedure, heterogeneity remained high regardless of the exclusion of any single study.

Figure 8 Withdrawal time for colonoscopy with vs without computer-aided detection assistance on first pass for the included studies. A: Forrest plot; B: Sensitivity analysis. CADe: Computer-aided detection.

Figure 9 Forrest plot showing withdrawal time for colonoscopy with vs without computer-aided detection assistance on second pass for the included studies. CADe: Computer-aided detection.

DISCUSSION

Previous meta-analyses have also reported favorable outcomes associated with AI-assisted colonoscopy. However, most studies have primarily focused on ADR, with AMR receiving only brief discussion and rarely being directly compared with ADR. The present analysis demonstrated a significant reduction in overall AMR with CADe assistance. However, its effectiveness was reduced in the ≥ 5 mm subgroup, which may be attributed to operator-dependent limitations during colonoscopy. Small and well-hidden adenomas are inherently difficult to detect with the naked eye, potentially explaining differences in CADe performance between advanced adenomas and sessile serrated adenomas. While advanced adenomas are typically ≥ 5 mm, routine colonoscopy generally achieves high detection accuracy. However, since this accuracy remains influenced by operator experience, the CADe system proved promising in mitigating this limitation by enhancing detection support[24].

The double-pass colonoscopy design used in the included studies might have influenced the observed ADR and APC, as the second examination inherently increases the likelihood of adenoma detection. To account for this effect, ADR from the first-pass examination was analyzed separately to provide a more accurate assessment of single-procedure ADR. Although first-pass ADR demonstrated statistically significant differences, regional variations in baseline ADR values and operator-related psychological factors may affect AI sensitivity in adenoma detection, making AMR a more reliable indicator of AI[25,26]. For instance, a study reported differences in adenoma detection sensitivity between Eastern and Western endoscopists (85.0% vs 75.8%)[27]. Additionally, even among endoscopists with an ADR value of 40%, adenomas have been found to be missed in approximately one-quarter of patients[28]. Despite its advantages, AMR may have limitations in fully evaluating the effectiveness of AI-assisted colonoscopy. Some studies suggest that missed adenomas may result from inadequate mucosal exposure, and improving mucosal visualization through enhanced techniques or mechanical assistance may further reduce AMR[21,29]. Additionally, AI-assisted colonoscopy may not significantly lower AMR in cases of inadequate bowel preparation[30], indicating that AMR remains susceptible to external factors. No significant differences were observed in withdrawal times between the CADe-assisted and routine colonoscopy groups. Given that sufficient withdrawal time is associated with more thorough examinations and improved procedural safety, this finding suggests that AI assistance does not compromise procedural quality[31].

AI-assisted colonoscopy has certain limitations. First, AI does not guarantee perfect detection, as its performance depends on the system’s training data and algorithms. While AI-generated findings often align with its predictive models, they may not fully reflect actual clinical scenarios, which may explain why AMR serves as a more sensitive indicator of colonoscopy performance[32]. Second, AI functions as a secondary reader, and its effectiveness remains dependent on the endoscopist’s skill and technique, as the system operates under the endoscopist’s guidance[33,34]. Despite these limitations, the benefits of AI-assisted colonoscopy outweigh the risks. AI supports endoscopists in maintaining consistent performance without fatigue, contributing to an overall improvement in ADR. However, an increase in adenoma detection may add to the medical burden on patients[35]. While AI-assisted colonoscopy raises individual healthcare costs, it has the potential to reduce CRC incidence, lower overall national healthcare expenditures, and alleviate pressure on medical insurance systems[36,37]. Moreover, AI-assisted colonoscopy has been shown to be cost-effective, with one study reporting an average savings of $57 per examination per patient in the United States[38]. In the long term, AI has the potential to improve colonoscopy by enhancing the speed and accuracy of lesion detection. Since CRC primarily develops from polyps or adenomas, identifying and removing these precancerous lesions represents a viable strategy for cancer prevention. However, due to time constraints, no studies have definitively confirmed the long-term impact of AI-assisted colonoscopy on CRC prevention. Nonetheless, detecting more precancerous lesions is likely to contribute to more effective prevention strategies.

This study has several limitations. First, the relatively small number of included patients may limit the generalizability of the findings and may not fully reflect colonoscopy performance across the broader population. Additionally, all included studies were conducted in Northern Hemisphere countries, with a predominance in Asia. The effectiveness of AI-assisted colonoscopy in Western countries and the Southern Hemisphere remains uncertain and requires further clinical validation. Second, with the exception of one study, all included trials employed a single-blind design. This may have increased awareness among endoscopists, potentially influencing outcomes and limiting the generalizability of the results. Third, all studies utilized a tandem design, involving two consecutive colonoscopies performed within a short time frame, which does not reflect routine clinical practice. While CADe-assisted colonoscopy has demonstrated potential benefits, its clinical applicability should be further assessed through large-scale prospective studies to validate its effectiveness in real-world settings.

In recent years, continuous advancements in medical technology have driven the evolution of AI systems, leading to the development of strategies such as the dual AI approach[39]. This approach involves the simultaneous application of multiple AI techniques or the integration of AI with other endoscopic technologies to enhance ADR. One study reported that when AI was combined with the Endocuff, a device attached to the distal end of the colonoscope to improve mucosal exposure, ADR increased by 4.9% and PDR improved by 3.0%. Even after the removal of Endocuff, AI-assisted colonoscopy remained superior to standard colonoscopy, with significantly higher ADR (53.8% vs 46.3%, P < 0.01) and PDR (74.0% vs 54.2%, P < 0.01). However, the decrease in ADR and PDR after Endocuff removal suggests that the combination of AI and mechanical enhancement provides a greater diagnostic advantage than AI alone[40]. Beyond hardware innovations, ongoing improvements in AI algorithms have further enhanced detection capabilities. A study found that the YOLOv5s + BiFPN model exhibited superior accuracy and recall compared to the standard YOLOv5 algorithm, highlighting the potential of advanced AI models in colonoscopy[41]. The use of real-time assistive technology in AI-assisted colonoscopy has also gained increasing attention. One study demonstrated that integrating real-time assistive systems, such as the CAQ system, with AI significantly improved ADR (30.6% vs 21.27%, P = 0.024)[42]. The combined application of AI has shown promise in addressing challenges associated with detecting extrinsic adenomas. Although AI is increasingly being implemented in clinical practice, colonoscopy procedures are typically performed by experienced endoscopists. Training in endoscopy requires extensive time and incurs significant costs. AI-assisted colonoscopy has been shown to enhance diagnostic accuracy among trainees. A retrospective study demonstrated that AI-assisted colonoscopy significantly improved the diagnostic performance of trainees, achieving an AMR comparable to that of experienced endoscopists[43]. This suggests that CADe-assisted colonoscopy may reduce training duration while facilitating a higher level of proficiency more rapidly. To date, CADe has been widely recognized for its ability to reduce human error and potentially lower the risk of colorectal lesions progressing to cancer. However, its implementation may also lead to an increased rate of unnecessary adenoma removal[44].

CONCLUSION

In conclusion, CADe-assisted colonoscopy offers significant advantages over standard colonoscopy, including improved ADR and a reduction in AMR. Additionally, it may alleviate the workload of endoscopists while enhancing procedural efficiency. Current evidence supports the clinical application of AI in colonoscopy, underscoring its potential to improve diagnostic accuracy and optimize workflow.