Colorectal cancer is the third cause of death and the fifth in frequency, worldwide, with a growing incidence in recent decades.1 Lower gastrointestinal endoscopy is one of the best screening strategies for colorectal cancer.2 A critical quality indicator of colonoscopies is the adenoma detection rate. A higher adenoma detection rate has been associated with a lower incidence of colorectal cancer.3 Nevertheless, reports in the literature describe a high miss rate, up to 30%, in colonic polyp detection.4

Artificial intelligence (AI) has garnered great interest in recent years in the field of gastroenterology, especially regarding colonic polyp detection during screening examinations. Different studies have demonstrated the benefit of AI systems in identifying raised lesions in the colon, with adequate sensitivity and specificity values.5–9 However, there is limited information on its actual application in daily clinical practice.

Therefore, the present study aimed to compare the polyp and adenoma detection rates between AI-assisted colonoscopy (AIAC) and control colonoscopy in a real-world setting.

A cross-sectional study was conducted at the gastroenterology service of the Clínica Angloamericana in Lima, Peru, applying the STROBE checklist for cross-sectional studies. Complete colonoscopies (obligatory cecal intubation), performed within the time frame of January 2021 and May 2024 by the same expert endoscopist (over 10 years of experience, with more than 1,000 annual procedures performed), were reviewed. Screening and diagnostic colonoscopies (patients with abdominal pain, chronic anemia, positive fecal occult blood test and/or chronic diarrhea) were included. Procedures with poor bowel preparation (Boston scale < 6 points), colonoscopies in patients with previous colon surgery, patients with a history of colon cancer, patients suspected of having hereditary gastrointestinal polyposis syndrome, and patients diagnosed with inflammatory bowel disease were excluded.

Bowel preparation was carried out with a split-dose regimen of 4 polyethylene glycol packets and 4 tablets of bisacodyl, 12 h before the colonoscopy. Colonoscopy was considered complete when cecal intubation was performed. The proximal colon was examined twice, utilizing the cap-assisted approach.

The polyp characteristics collected were location, morphology (Paris classification), size, and histology. The adenoma detection rate was defined as the number of colonoscopies with adenomatous polyps (including serrated adenomas) divided by the total number of colonoscopies performed in the same period. The polyp detection rate was calculated by the number of colonoscopies with at least one polyp divided by the total number of colonoscopies performed in the same period. The large intestine was anatomically divided into the proximal colon (cecum, ascending colon, and transverse colon) and the distal colon (descending colon, sigmoid colon and rectum). Three pathologists from the pathologic anatomy service at the Clínica Angloamericana were responsible for the histology review. Resected lesions < 10 mm were placed in jars labeled by segment (proximal or distal colon) and only the specimens > 10 mm were placed in separate jars.

All the procedures reviewed were performed using Eluxeo® EC-760ZP-V/L (Fujifilm, Japan) equipment. After March 2022, the CAD EYE (Fujifilm, Japan) AI system, designed for white-light colonoscopy, was used. The system’s interface highlights the polyp through a colored square around the raised lesion, emitting a beeping sound during detection. Finally, a visual assist circle lights up toward the polyp. The procedures were divided into a control group (performed from January 2021 to February 2022) and the AIAC group (performed from March 2022 to May 2024).

A total of 1,251 colonoscopies that met the selection criteria were evaluated. The median patient age was 58 years (range: 49-68 years), with a higher proportion of women (56%). A screening colonoscopy was the main indication for the procedure (68%). The overall adenoma detection rate for the study was 37%, with an adenoma detection rate of 37% for men and 36% for women.

In the per-patient analysis, the polyp detection rate with AIAC was significantly higher than that of the control colonoscopies (58% vs 52%, p < 0.05). AIAC detected at least one adenoma in 321 of 828 patients, compared with 140 of 423 patients in the control group, corresponding to adenoma detection rates of 39% and 33%, non-significantly higher in the AIAC group (Table 1). Even though there were age differences between the two groups, there was no significant variation between the adenoma detection rates of the control colonoscopies and the AIACs, between age groups, after 50 years of age (Fig. 1). Regarding serrated adenomas, AIAC had a non-significant higher rate, compared with control colonoscopy (6% vs 5%, respectively).

Figure 1.

Age group analysis.

AIAC: artificial intelligence-assisted colonoscopy; CC: control colonoscopy.

In the colonoscopy indication sub-analysis, the use of AI increased the detection of polyps, adenomas, and serrated adenomas, with no statistical significance, for the two indications (Table 2). Likewise, in the AIAC group, the adenoma detection rate for the years 2022, 2023, and 2024 remained stable for the screening indication (44%, 40%, and 41%, respectively) and the diagnostic indication (31%, 35%, and 28%, respectively) (Fig. 2).

Figure 2.

Analysis of the adenoma detection rate by indication for colonoscopy and year.

Table 3 shows the polyp analysis (n = 1,135). Of all the samples, 19 were classified as having normal mucosa (1.7% false positives): 10 of 336 samples (2.9%) for the control group and 9 of 799 samples (1.1%) for the AIAC group, and the difference between the two groups was statistically significant (p = 0.039). There was a non-significant higher number of tubular adenomas with low-grade dysplasia in the AIAC group, compared with the control group.

Fig. 3 summarizes the adenoma characteristic analysis. AIAC significantly increased the detection of polypoid adenomas (OR: 1.28; 95% CI 1.04-1.59), < 10 mm (OR: 1.41; 95% CI 1.14-1.74), and located in the proximal colon (OR: 1.31; 95% CI 1.05-1.65), compared with control colonoscopy.

Figure 3.

Simple logistic regression analysis based on polyp and adenoma characteristics.

AIAC: artificial intelligence-assisted colonoscopy.

To the best of our knowledge, this is the first study conducted in Peru that describes the performance of AIAC in daily clinical practice. The primary aim was to compare the adenoma detection rate between conventional colonoscopy and AIAC. The use of AI showed an increase in the adenoma detection rate (39% vs 33%), mainly for adenomas < 10 mm and in the proximal colon, that was not statistically significant.

Different studies have evaluated diagnostic yield in the identification of polyps and adenomas. In an analysis of 5 randomized clinical trials with 4,354 patients, conducted by Hassan et al.,10 there was a 70% relative increase in adenoma per colonoscopy and a 44% relative increase in the adenoma detection rate with AIAC, compared with conventional colonoscopy, without affecting the efficiency of the endoscopic examination. This result aligns with our findings on AIAC performance, with a large number of procedures performed by a single experienced endoscopist, although it did not reach statistical significance. However, in their retrospective, observational study that included 4,414 colonoscopies, Levy et al.11 reported that AI applied in clinical practice did not increase the polyp and adenoma detection rates, unlike the increase shown with our results. However, they described a shorter procedure time in the colonoscopies with AI and did not report on bowel preparation, factors that could have an influence on the detection of raised lesions.

In our study, AIAC significantly increased the likelihood of detecting polypoid adenomas smaller than 10 mm and located in the proximal colon, compared with control colonoscopy. Those findings align with the results described by Repici et al.,9 who observed an increase in flat adenoma detection and adenomas smaller than 10 mm, through AI use performed by endoscopists with limited experience. Wallace et al.,12 whose randomized tandem study aimed to assess the adenoma miss rate, found that AI might enhance the detection of adenomas smaller than 10 mm in the distal colon. A possible explanation could be the level of mucosal exposure, which may increase the visual field and the opportunity to identify raised lesions. Future randomized studies are needed to evaluate the potential synergy between AI systems and mucosal exposure devices, such as Endocuff or cap attachments.

Different studies report an increase in the incidence of interval tumors at the level of the proximal colon, with colorectal cancer appearing between one colonoscopy and the next recommended procedure.13 Thus, there are strategies for increasing the adenoma detection rate in that specific region of the colon. Chandran et al.14 reported a significant decrease in missed adenomas by evaluating the proximal colon through retroflexion (26.4% vs 24.6%, p < 0.001). Other techniques described for increasing adenoma detection yield are the double evaluation of the right colon and the use of mucosal exposure accessories, such as the cap attachment,15 techniques applied in our study. Based on our results, we recommend AIAC in clinical practice for increasing raised lesion detection, consequently reducing the incidence of colon cancer.

Serrated polyps are responsible for 30% of tumors of the colon.16 Different international guidelines recommend a serrated adenoma detection rate above 7%, and ideally above 10%.3 Similarly, Ellison et al.17 demonstrated that the double examination of the right colon increased the serrated polyp rate by 10%, in the hands of experienced endoscopists. Even though we found no significant increase in our serrated adenoma detection results, the use of AI may increase the colonoscopic identification of serrated adenomas, not only when performed by trainees, but also by qualified professionals working at high-volume centers.

A strength of the present study is the large number of procedures performed. The use of AI in our study (Table 2 and Fig. 2) enabled a higher adenoma detection rate than that suggested for colorectal cancer screening programs,18,19 consistent with previously published data, in which AI-assisted detection systems reduce the adenoma miss rate when performed by experienced endoscopists,20 as well as by trainees.9 This advantage is very important in large-volume centers and in training centers for endoscopists. Another strength of the present study is that the endoscopies were performed by the same experienced endoscopist (with more than 10 years of experience), thus reducing interobserver bias and increasing the validity of our results.

Of the study’s limitations, first, there was a significant age difference between the AIAC patients and the control colonoscopy patients. However, in the age range sub-analysis, there were no significant differences in the adenoma detection rates in patients above 50 years of age, which is the group at higher risk for adenomas, thus reducing selection bias. Second, we had no register of colonoscopy withdrawal time, an important quality indicator. However, there were no significant differences in total procedure time (cecal intubation time and withdrawal time) between the standard colonoscopies and the AIACs. Importantly, AI use does not affect withdrawal time,21 which indirectly correlated with the absence of differences in the total colonoscopy time between the two groups in our study. Third, there was an insufficient number of positive immunochemical fecal occult blood tests in the colonoscopies for making a sub-analysis. Due to the retrospective design of the study, there were no data on family histories of colon cancer, nor were patients with previous colonoscopies who had a history of polyps or adenomas, excluded. To validate the findings of our study, we recommend conducting new trials with the participation of more than one endoscopist.

In conclusion, AIAC increased the polyp detection rates, particularly of polypoid lesions under 10 mm in size and located in the proximal colon. Randomized trials should be conducted that evaluate the probable benefit of mucosal exposure devices and AI-assisted detection systems in daily clinical practice.

The protocol of the present study was approved by the ethics committee of the Clínica Angloamericana (no. CIEI-CAA-067/2024) and was conducted following the recommendations of the Declaration of Helsinki (Fortaleza, Brazil, 2013). Patient data were encrypted to prevent identification, stored in the clinic’s digital media, and managed exclusively by the authors of the study.

Study concept and design: RZH and JAC; data acquisition: BG and JAC; data analysis and interpretation: BG, RZH, and JAC; transcription and critical review of the intellectual content of the manuscript: BG, RZH, and JAC; definitive approval of the final version: BG, RZH, and JAC.

No specific grants were received from public sector agencies, the business sector, or non-profit organizations in relation to this study.