Impact of Endocuff addition to real-time computer-aided detection of colorectal neoplasia in a randomised tandem colonoscopy trial

Abstract

BACKGROUND

Risk of post colonoscopy colorectal cancer is related to adenoma miss rate (AMR) during colonoscopy. Artificial intelligence (AI) and Endocuff Vision are tools both used to increase adenoma detection rate (ADR).

AIM

To assess whether the combination of AI and Endocuff vision, compared to using AI alone, increases the ADR.

METHODS

This is a single-center randomized, tandem colonoscopy trial. Patients with a Boston Bowel Preparation score ≥ 6 with an indication of colorectal cancer screening (78, 91.8%) or polyp surveillance (7, 8.2%) were included from the study.

RESULTS

Eighty-five patients were included in total (male: 51; mean age: 63 ± 8 years old). In 39 patients, the initial colonoscopy was performed using AI alone, while in 45 patients, it was carried out with a combination of AI and Endocuff. Colonoscopies without Endocuff were associated with a numerically higher ADR (19/39, 48.7% vs 14/45, 31.1%, P = 0.107) and an increased number of polyps detected per procedure (1.7 ± 2.5 vs 1.2 ± 1.4, P = 0.272). During tandem colonoscopy, an additional 0.4 ± 0.8 polyps per examination (polyp miss rate = 0.08 ± 0.15), along with 0.3 ± 0.7 adenomas (AMR = 0.10 ± 0.25), were identified. Adding Endocuff to AI during tandem colonoscopy did not provide any benefit in AMR or polyp miss rate when compared to initial AI combined with Endocuff-assisted endoscopy followed by tandem AI-only procedures (0.11 ± 0.24 vs 0.92 ± 0.27, P = 0.727; and 0.10 ± 0.15 vs 0.07 ± 0.14, P = 0.415, respectively).

CONCLUSION

Adding Endocuff to AI-assisted colonoscopy, according to our results, does not lead to increased adenoma detection.

Key Words: Adenoma detection rate; Adenoma missing rate; Artificial intelligence; Endocuff Vision; Screening colonoscopy

Core Tip: Adenoma detection rate (ADR) is associated with a decreased risk of developing interval colorectal cancer, whereas the adenoma miss rate during colonoscopy is related to the risk of post-colonoscopy colorectal cancer. Artificial intelligence (AI) and Endocuff Vision, have independently shown to enhance ADR. However, in our prospective tandem study assessing the combined use of Endocuff Vision and AI, we did not observe validation of previously published data indicating that adding Endocuff to AI-assisted colonoscopy improves overall polyp and ADRs.

INTRODUCTION

Colonoscopy is recognized as the leading screening tool for colorectal cancer (CRC)[1,2]. It offers the highest combined sensitivity and specificity for detecting CRC, and it also allows for the identification and removal of precancerous lesions such as adenomas, thereby playing an important preventive role[1,2]. The adenoma detection rate (ADR) - the proportion of screening colonoscopies in which an adenoma is found - is the most widely used, studied, and validated metric for assessing the quality of screening colonoscopy, as a higher ADR correlates with a reduced risk of developing interval CRC[3]. A more rigorous measure of screening effectiveness is the adenoma miss rate (AMR), which is directly associated with the incidence of interval cancers[4].

However, colonoscopy is still not a perfect screening modality, as even experienced endoscopists may miss up to one in four adenomas[5]. Over the years, numerous strategies have been implemented to improve the ADR, including modifications of basic maneuvers, advanced insertion techniques, technological innovations that enhance image visualization, auxiliary devices designed to expand the visual field, and, more recently, the incorporation of artificial intelligence (AI).

Incomplete visualization of blind spots - most notably the regions obscured by colonic folds - remains a major contributor to missed polyps during colonoscopy. To enhance mucosal inspection, multiple distal attachment technologies have been developed, among them permanently integrated systems like the G-Eye and removable devices such as the Endocuff, EndoRing, and transparent cap. Regarding Endocuff-assisted colonoscopy (EAC), randomized controlled trials - including those conducted within various national bowel cancer screening programs - have not consistently demonstrated a significant increase in ADR[6-10]. Despite this, meta-analytic data suggest that EAC yields a significant enhancement in ADR, particularly among endoscopists with baseline rates below 30%, for whom an absolute improvement of 9.4% has been reported (P = 0.03)[11-13]. Similar to other auxiliary devices, EAC has not shown consistent or substantial improvements in AMR.

With the incorporation of AI into endoscopy, computer-aided detection (CADe) systems have emerged, using deep learning models to facilitate real-time recognition of lesions during colonoscopy. The United States Food and Drug Administration approved the first CADe system in 2020[14], and several additional systems have been commercially introduced since then. CADe technology holds significant potential to reduce operator variability and improve ADR. Multiple meta-analyses, both past and recent, indicate that AI-assisted colonoscopy increases ADR by approximately 8% to 20%[15-17], with most evaluated CADe systems demonstrating similar performance risk ratios ranging from 1.14 to 1.27[17]. Furthermore, CADe has also contributed to improvements in AMR[18]. The primary aim of our study was to evaluate if the combination of AI and Endocuff Vision compared to using AI alone, increases the effectiveness of screening colonoscopy.

MATERIALS AND METHODS

This is a single-center randomized, tandem colonoscopy trial in which each patient underwent back-to-back colonoscopies performed by the same endoscopists on the same day. The sequence of the procedures was randomized: The first colonoscopy was performed using AI with Endocuff Vision and the second using AI alone and vice versa. All procedures were performed by 2 experienced endoscopists using high definition colonoscopes (CF-H185 L, Olympus, Japan) and the GI Genius (Medtronic Minneapolis, MN, United States) CADe system (Figure 1). The use of antispasmodics and patient position changes during both insertion and withdrawal were recorded.

Figure 1 Image representation of Endocuff Vision and GI Genius computer-aided detection system. A: Endocuff Vision; B: Computer-aided detection system. Flat lesion detected by computer-aided detection system (green box).

Major metrics assessed

Major metrics assessed: (1) AMR: The number of adenomas found in the second colonoscopy by the total number of adenomas detected in both colonoscopies combined; (2) ADR: The proportion of screening colonoscopies in which an adenoma is found; (3) Polyp miss rate (PMR): The number of polyps found in the second colonoscopy by the total number of polyps detected in both colonoscopies combined; and (4) Polyp detection rate (PDR): The proportion of screening colonoscopies in which a polyp is found.

Study inclusion criteria

Patients referred for colonoscopy for: (1) Screening; (2) Surveillance post-polypectomy; or (3) Diagnostic reasons.

Study exclusion criteria

Study exclusion criteria: (1) Patients with a total Bristol Bowel Preparation score (BPPS) < 6 or BPPS = 0 in any segment of the colon; (2) A history of or newly diagnosed polyposis syndrome; (3) Known colonic stricture; (4) Known severe diverticular disease; (5) Active colitis; and (6) Failure to discontinue antiplatelet/anticoagulant medications. Initially 85 patients were enrolled; however, three patients were excluded: One due to inability to tolerate the procedure; one due to newly diagnosed serrated polyposis syndrome; one due to BPPS < 6. Patients were subsequently divided into two groups: (1) Group CAD and E (Endocuff): Underwent colonoscopy first with AI and Endocuff Vision, followed by AI alone; and (2) Group CAD: Underwent colonoscopy first with AI alone, followed by AI and Endocuff Vision. A study flowchart is provided in Figure 2.

Figure 2 Study flowchart. CAD: Computer-aided detection.

Statistical analysis

Data were analyzed using SPSS v27 (IBM SPSS Statistics, Chicago, IL, United States). Variables were expressed as frequencies, means ± SD, or medians with interquartile ranges, as appropriate. Continuous variables were compared using t-test for normally distributed data or the Mann-Whitney U test for non-normally distributed data. Categorical variables were compared using the χ² or Fisher’s exact test, as appropriate. Associations between variables were assessed with Spearman’s correlation coefficient. Multivariate linear and logistic regression models were used to identify independent predictors. Only variables with P < 0.10 in univariate analysis were included in the multivariate models. Two-sided P values < 0.05 were considered statistically significant. A formal ethical approval was received by the Research and Ethics Committee of Aretaieion University Hospital (approval No. 489/14-03-2023).

RESULTS

General epidemiological data

In total, 85 patients were included in the study (51 male and 34 female participants; mean age 63 ± 8 years). Of these, 39 underwent initial colonoscopy with AI assistance alone, whereas 45 were examined using a combination of AI and Endocuff. The predominant indication for colonoscopy was CRC screening, representing 91.8% of all cases. Baseline characteristics are summarized in Table 1.

Table 1 Baseline characteristics of all included patients, n (%)/mean ± SD.

Characteristic	Value
Age (years)	62.58 ± 7.587
Male sex	50 (58.8)
Indication	85 (100)
Screening	75 (88,2)
Surveillance	7 (8.2)
Insertion time (minutes)	5.84 ± 2.52
Bristol Bowel Preparation score	8.57 ± 1.011
Position change	79 (96.3)
Butylscopolamine	39 (47.5)
Withdrawal time (minutes)	7.2 ± 2.140

No significant differences were observed between the groups with respect to bowel preparation quality, demographic characteristics, or indication for colonoscopy. Withdrawal time did not differ significantly either (7.2 ± 2.5 minutes with Endocuff vs 7.2 ± 1.7 minutes without; P = 0.887). Notably, moderate to severe procedural discomfort was more frequently reported in the Endocuff group, with the difference nearing statistical significance (4/43; 9.3% vs 0/39; P = 0.071). A total of 120 polyps were detected including 53 adenomas, 68 diminutive polyps and 18 advanced adenomas (defined as serrated polyps > 1 cm, adenomas with low grade dysplasia > 1 cm and adenomas and adenomas with high grade dysplasia). A comparison of baseline characteristics between the two groups is presented in Table 2.

Table 2 Comparison of baseline characteristics between the two groups, n (%)/mean ± SD.

Characteristic	CAD (n = 39)	CAD and E (n = 43)	P value
Age (years)	63 ± 8	62 ± 7	0.453
Sex (male/female)	20/19	29/14	0.117
Bristol Bowel Preparation score	8.5 ± 0.9	8.7 ± 0.5	0.137
Insertion time (minutes)	5.82 ± 2.73	5.86 ± 2.34	0.943
Position change	37 (94.8)	42 (97.6)	0.456
Butylscopolamine	9 (30)	30 (69.7)	< 0.001
Withdrawal (minutes)	7.2 ± 1.7	7.2 ± 2.5	0.887
Discomfort (≥ 1)	16 (41)	19 (44.1)	0.474

CAD: Computer-aided detection; CAD and E: Computer-aided detection and Endocuff.

Butylscopolamine use

Among all patients the use of butylscopolamine did not affect either the PMR or ADR. The PMR was 0.081 ± 0.158 with butylscopolamine vs 0.083 ± 0.141 without it (P = 0.959), and the AMR was 0.130 ± 0.314 with butylscopolamine vs 0.079 ± 0.185 without it, (P = 0.362).

PDR and PMR among the two groups

The PDR did not significantly differ between procedures performed with Endocuff (25/39, 64.1%) and those without it (25/43, 58.1%, P = 0.373), although the number of polyps detected per colonoscopy was numerically higher without Endocuff (1.7 ± 2.5 vs 1.2 ± 1.4, P = 0.272) (Figure 3A). No significant differences in PDR were observed in the left or right colon. Using Endocuff, PDRs were 16.3% in the left colon and 44.2% in the right colon, whereas the corresponding rates without Endocuff were 28.2% and 51.3% (P = 0.150 and P = 0.307, respectively). During tandem colonoscopy, an additional 0.4 ± 0.8 polyps per examination were detected resulting in a PMR of 0.08 ± 0.15. The addition of Endocuff during tandem colonoscopy (among patients initially examined with AI alone), did not significantly affect the PMR (0.10 ± 0.15 vs 0.07 ± 0.14; P = 0.415).

Figure 3 Polyp detection rate and adenoma detection rate in the total cohort and among study groups. A: Polyp detection rate; B: Adenoma detection rate. CAD: Computer-aided detection; CAD and E: Computer-aided detection and Endocuff.

When analyzing only patients who did not receive butylscopolamine, no significant differences in PMR were observed between the two groups (CAD and E: 0.256 ± 0.153 vs CAD: 0.108 ± 0.153, P = 0.088). Similarly, among patients who did receive butylscopolamine, PMR remained comparable (CAD and E: 0.092 ± 0.165 vs CAD: 0.044 ± 0.13333; P = 0.434).

ADR and AMR among the two groups

The ADR was numerically higher without Endocuff (19/39, 48.7%) compared to with Endocuff (14/43, 32.6%), although this difference did not reach statistical significance (P = 0.107) (Figure 3B). There were no significant differences in ADR in the left or right colon between the two groups. With Endocuff detection rates were 14.0% (left) and 37.2% (right) vs 20.5% (left) and 48.7% (right) without Endocuff (P = 0.559 and P = 0.372, respectively).

Tandem colonoscopy identified an additional 0.3 ± 0.7 adenomas per procedure, resulting in an AMR of 0.10 ± 0.25. Among patients initially examined with AI alone, the addition of Endocuff during the tandem examination did not yield a significantly higher AMR compared with those initially examined using AI + Endocuff and subsequently with AI alone (0.11 ± 0.24 vs 0.92 ± 0.27; P = 0.727).

When analyzing only patients who did not receive butylscopolamine, no significant difference in AMR was observed (CAD and E: 0.001 vs CAD: 0.128 ± 0.210; P = 0.660). Similarly, among patients who did receive butylscopolamine, AMR remained comparable (CAD and E: 0.136 ± 0.314 vs CAD: 0.111 ± 0.333; P = 0.838). No adverse events were reported during the study.

DISCUSSION

Recent interest has been raised regarding the combination of novel techniques and devices to improve ADR. In this randomized tandem colonoscopy trial, we evaluated the additive effect of Endocuff Vision combined with AI-assisted colonoscopy on ADR and AMR. Our data did not demonstrate a significant increase in ADR or a reduction in AMR with the combination of Endocuff Vision and AI compared to AI alone. The numerical trend toward higher ADR and PDR without Endocuff was unexpected but not statistically significant.

Previous randomized controlled trials showed a significant increase in ADR with the combination of Endocuff and AI[19,20]. Contrary to expectations based on these studies, our study failed to demonstrate comparable results. Nevertheless, our outcomes are in agreement with some randomized controlled trials and meta-analyses suggesting that Endocuff may not provide a substantial benefit to ADR in populations, where baseline detection rates are already high[9,21-23]. In particular, a higher anticipated benefit of Endocuff has been reported for low-performing endoscopists with lower baseline ADRs, whereas the improvement might be only marginal among expert operators[24].

Another explanation for the lack of additive benefit from Endocuff could be a potential ceiling effect with AI. Although AI devices provide information solely about lesions within the field of view, it is questionable whether their capabilities in mucosal visualization and lesion detection diminish the incremental advantage offered by Endocuff’s mechanical flattening of the folds. When sub-analyses were performed in the left and right colon, the combination of Endocuff with AI failed to demonstrate an additional gain in the left colon - where most blind spots are traditionally observed - further reinforcing the above hypothesis.

In our study, dynamic position changing was recorded but not included in the initial randomization. An additional gain from combining dynamic position changes with AI warrants further investigation. Moreover, the AMR and PMR during tandem colonoscopy were low and comparable between groups, reinforcing the finding that Endocuff did not improve detection beyond AI. We acknowledge the operator bias inherent in tandem colonoscopy studies; however, no significant differences in procedure-related parameters such as withdrawal time or bowel preparation quality were observed between groups, supporting the notion that the findings are unlikely to be due to procedural confounders. This is a clinically relevant outcome since AMR has a direct link to interval CRC risk[4].

There are limitations to our study. It is a single-center study with a relatively small sample size, which may reduce the power to detect small differences in ADR or AMR. Moreover, both endoscopists performing the procedures were highly experienced. We did not include less experienced or low-detector endoscopists, so it remains uncertain whether AI combined with Endocuff would be equally beneficial in these groups. Finally, we cannot exclude the possibility of psychological bias, as the operator was aware of the randomization.

CONCLUSION

Our study failed to demonstrate the superiority of the combination of Endocuff Vision and AI over AI alone in increasing ADR or reducing AMR. Currently, several large multicenter trials are underway to evaluate the combined efficacy of emerging technologies across different clinical settings and operator experience levels. There remains limited data on the synergistic effect of combining simple procedural techniques with technological innovations. Cost-effectiveness analyses would also be valuable to assess whether integrating these technologies justifies the additional expense.