Linked color imaging versus artificial intelligence-assisted linked color imaging for neoplasia detection in the colorectum: a randomized trial in Brazil.

Graphical abstract

INTRODUCTION
The removal of colorectal adenomas by colonoscopy reduces the morbidity and mortality of colorectal cancer (CRC).1 This is attributed to the adenoma–carcinoma sequence, which is considered a major pathway in the development of CRC.2 However, approximately 15% to 30% of CRCs develop via the serrated pathway.3,4 Therefore, several guidelines recommend the detection and resection of both adenomas and sessile serrated lesions (SSLs) during colonoscopy for CRC prevention.5,6 Nonetheless, a systematic review has revealed miss rates of 26% for adenomas, 9% for advanced adenomas (AAs), and 27% for serrated polyps.7 Reported risk factors for missed polyps include poor bowel preparation, right-sided location, nonpolypoid morphology, small size, and SSLs.8
Various image-enhanced endoscopy (IEE) systems have been developed to reduce polyp miss and improve adenoma detection rates (ADRs). The use of a laser endoscope (LASEREO; Fujifilm Co.) allows for blue laser imaging (BLI) and linked color imaging (LCI) as a form of narrow-band light observation.9 Additionally, a light-emitting diode (LED) endoscope system (ELUXEO; Fujifilm Co.) enables BLI and LCI with multi-light technology.10 Many randomized controlled trials (RCTs) have demonstrated the efficacy of polyp detection, including adenomas and SSLs, with LCI for both laser and LED systems.11-14
Regarding artificial intelligence (AI) in colonoscopic practice, some meta-analyses have shown its effectiveness in polyp detection, indicating a significantly higher ADR for endoscopists using AI than for those not using AI.15,16 A computer-aided detection (CAD) platform based on AI deep learning technology, CAD EYE (Fujifilm Co.), has enabled endoscopists to detect colorectal polyps using white light imaging (WLI) and LCI, and to diagnose polyp characteristics using BLI. Several studies have demonstrated the efficacy of lesion detection and diagnosis with CAD EYE using recorded images and videos.17-19 Additionally, recent clinical trials have shown the effectiveness of CAD EYE with WLI in real cases.20,21 A recent study has revealed that the use of CAD EYE significantly improved ADR compared with total colonoscopy without CAD EYE.22 In this study, we compared neoplasia detection rate (NDR) between LCI and AI-assisted LCI colonoscopy.

METHODS

Study design
We performed a single-center, prospective, randomized trial using a computer-generated list of random numbers and concealed the allocation sequence from those involved using sealed envelopes. A nurse opened the envelopes before the start of colonoscope withdrawal. This study was registered with the Brazilian Protocol Registry/REBEC (http://www.ensaiosclinicos.gov.br). The universal trial number is U1111-1320-2350.

Patients
Consecutive patients referred for colonoscopy in our department from October 2023 to March 2024 were recruited to participate in the study. Eligible participants were all patients aged ≥18 years undergoing colonoscopy for screening, surveillance, and symptoms (abdominal pain, diarrhea, constipation, abdominal distension, or tenesmus) who consented to participate in the study. Participants were blinded to each group. The Boston bowel preparation scale (BBPS) was used to assess colon preparation. Patients were excluded if they were younger than 18 years; had inadequate bowel preparation (BBPS score <6); had advanced CRC; scheduled for adenoma removal (previous diagnosis of adenoma); had a history of colorectal resection; had a diagnosis of familial adenomatous polyposis, inflammatory bowel disease, acute lower gastrointestinal bleeding, or actinic rectitis; had an incomplete colonoscopy; or were receiving antithrombotic therapy (clopidogrel, prasugrel, or ticagrelor, not aspirin). The patients were randomized in a 1:1 ratio into the LCI and AI-assisted LCI groups. Patients were divided by sex (male and female) and age (<50 years and ≥50 years) for analysis.

Endoscopic procedures
An endoscopist with experience in IEE since 2007 and some experience in AI (100–120 procedures) performed all colonoscopies using a high-definition colonoscope (EC-760ZP-V/L; Fujifilm Co.) with an ELUXEO 7000 system, which contains four LEDs that allow the use of IEE in the BLI, BLI-bright, and LCI modes, in addition to WLI. In LCI, the color contrast is enhanced to depict red and white colors more brightly, thus facilitating visualization and discrimination between lesions and normal mucosa. The CAD EYE system was used to demarcate the detected lesion by drawing a bounding box around the lesion, in addition to emitting a sound when detection occurred.
Bowel preparation included a fiber-free, clear-liquid diet for 1 day, with bowel cleansing by ingesting 1 L of 10% mannitol solution on the day of the colonoscopy. All patients received intravenous midazolam and fentanyl for sedation (conscious sedation). WLI mode alone was used to insert the colonoscope in all cases. The LCI mode was used during withdrawal in all patients with AI-assistance in one group.
The lesion location, size, morphology, and histology were also assessed. The location was divided into the right colonic segment (from the transverse colon to the cecum) and left colonic segment (from the rectum to the descending colon). An open biopsy forceps was used as a guide to measure lesion size, which was divided into three groups: ≤5 mm, 6–9 mm, and ≥10 mm. Lesion morphology was described as polypoid or nonpolypoid based on the Paris classification.23 All detected lesions were removed only during colonoscopy withdrawal. The withdrawal time was >6 minutes for all colonoscopies. A second inspection of the proximal colon and retroflexion of the right colon were not performed.
Specimens were fixed in 10% formalin and routinely processed for histological examination. The World Health Organization classification of colorectal tumors was used for histology.24 Neoplasia was defined as any adenomatous lesion or SSL. AA was defined as a lesion ≥ 10 mm in size, a villous component, or high-grade dysplasia. ADR was defined as the percentage of colonoscopies in which at least one adenoma was detected. Similar definitions were used for the SSL detection rate (SDR), NDR, and advanced adenoma detection rate (AADR). Each lesion was individually identified and stored for patients with multiple lesions. Both pathologists and patients were blinded to the randomization group.

Statistical analysis
Data were analyzed using Stata ver. 18.0 (StataCorp.). For the descriptive analysis, categorical variables are expressed as absolute and relative frequencies, and numerical variables are expressed as means and standard deviations (SDs). Differences in proportions were analyzed using Fisher’s exact test, whereas differences in mean values were analyzed using the Mann-Whitney U-test (when a parametric test was not suitable) or analysis of variance. The required sample size to detect a difference of 12 percentage points between the groups (LCI and LCI+AI) with an average of 50% in the ADR, considering a power of 80% and confidence limit of 95%, was 570 lesions (ratio 1:1). The analyses were stratified by group (LCI and LCI+AI) and treatment indications (screening, surveillance, and symptoms). To analyze the occurrence of adenomas, we ran a Poisson regression, with adjustment for sex and age, and considered the cluster effect (number of lesions versus number of patients). A two-sided 5% significance level was adopted for two-tailed tests. When appropriate, 95% confidence intervals (95% CIs) were used to indicate precision.

Ethical statements
The research ethics committee of Santa Casa de Caridade Hospital (processo 58/2024) approved this trial, which followed the tenets of the Declaration of Helsinki. Written informed consent was obtained from all study participants.

RESULTS
A total of 731 patients were randomized for examination by LCI (LCI group) or AI-assisted LCI (LCI+AI group) (Fig. 1). Of the patients, 109 were excluded (Fig. 2). A total of 622 patients were included in this study, and 779 polyps were detected in 416 patients.
Patient characteristics are described in Table 1. The mean patient age was 59 years; 475 (76.4%) patients were ≥50 years. Most patients were women (69.9%). The mean lesion size was 3.7 mm (SD=2.3). The overall polyp detection rate (PDR) was 66.9%. The PDR was 65.1% in the LCI group (n=304) and 68.6% in the LCI+AI group (n=318, p=0.39). The mean number of polyps per patient was 1.2 vs. 1.3 (p=0.14), respectively.
The adenoma characteristics are shown in Table 2. A total of 555 adenomas (tubular and tubulovillous) were diagnosed in 311 patients (259 adenomas in the LCI group vs. 296 in the LCI+AI group, p=0.63) with a mean age of 64 years, where 507 adenomas (91.4%) were detected in patients aged ≥50 years; 59.1% were women, with no difference between the groups (p=0.30). The mean adenoma size was 3.5 mm (SD=2.1). The overall ADR was 50.4% (95% CI, 46.2%–54.5%).
The ADR was 48.0% (95% CI, 42.1%–53.9%) in the LCI group and 52.6% (95% CI, 46.8%–58.4%) in the LCI+AI group (p=0.32). When screening, surveillance, and symptomatic patients were analyzed separately, the ADR was 56.6% (95% CI, 46.6%–66.5%) vs. 62.8% (95% CI, 53.8%–71.9%) (p=0.40), 48.6% (95% CI, 39.1%–58.2%) vs. 51.6% (95% CI, 42.7%–60.5%) (p=0.69), and 35.2% (95% CI, 23.8%–46.6%) vs. 33.3% (95% CI, 20.4%–46.3%) (p=0.85), respectively. The symptomatic subgroup did have fewer patients aged ≥50 years than the screening and surveillance subgroups in both the LCI group (55.7%, 75.0%, and 89.7%; p<0.001) and the LCI+AI group (51.6%, 80.5%, and 85.7%; p<0.01). The mean number of adenomas per patient (MAP) was 0.9 (0.9 in the LCI group vs. 0.9 in the LCI+AI group, p=0.19). The AADR was 5.8% (5.9% [95% CI, 3.3%–8.6%] in the LCI group vs. 5.7% [95% CI, 3.1%–8.2%] in the LCI+AI group, p=1.0). Regarding morphology, 66.7% of adenomas were nonpolypoid (62.5% in the LCI group vs. 70.3% in the LCI+AI group, p=0.06). Adenomas were more commonly located in the right colon (58.4%) and were more frequent in the LCI+AI group (52.1% vs. 63.8%, p=0.01).
A comparative description of the detection rates is provided in Table 3. The proportion of adenomas among the diagnosed polyps was higher in men than in women (85.6% vs. 76.8%, p=0.05) in screening colonoscopy patients (85.4% vs. 77.2% vs. 72.7%, p=0.05), in the case of polypoid morphology than of flat lesions (90.8% vs. 76.6%, p<0.01), and in the right colon (87.1% vs. 69.6%, p<0.01). A total of 62 SSLs were detected in 52 patients (28 in the LCI group vs. 34 in the LCI+AI group), with no difference between men and women. The SDR was 8.4% (8.2% [95% CI, 5.1%–11.3%] in the LCI group vs. 8.5% [95% CI, 5.4%–11.6%] in the LCI+AI group, p=1.0). SSLs were more commonly detected in patients aged <50 years than in those aged ≥50 years (27.3% vs. 9.7%, p<0.01), with statistical significance in the LCI group (40.0% vs. 7.7%, p<0.01), with no significance in the LCI+AI group (16.7% vs. 11.5%, p=0.41), and marginal significance comparing the LCI group and LCI+AI group in patients aged <50 years (40.0% vs. 16.7%, p=0.05). No differences in the morphology or location were observed between the groups. Regarding size, the lesions were predominantly >5 mm (p<0.01) and were unrelated to location in both groups. A total of 617 neoplastic lesions (79.2%) were detected in 363 patients: 287 in the LCI group and 330 in the LCI+AI group. The NDR was 54.5% (52.3% [95% CI, 46.7%–58.0%] in the LCI group vs. 56.6% [95% CI, 51.1%–62.1%] in the LCI+AI group, p=0.30). Cecal intubation and withdrawal times were 3.8 (95% CI, 3.5–4.0) vs. 3.9 (95% CI, 3.6–4.1) minutes (p=0.71) and 11.8 (95% CI, 11.4–12.3) vs. 11.8 (95% CI, 11.3–12.2) minutes (p=0.93), respectively.

DISCUSSION
ADR is considered the main quality indicator for colonoscopy and should be ≥35%.25 Lee et al.26 evaluated the quality of colonoscopy withdrawal technique and variability in ADR, dividing the endoscopists into three groups: low ADR (ADR <21%), moderate ADR (ADR, 21%–42%), and high ADR (ADR >42%).
Improved visualization affects the detection of CRC precursor lesions, and colonoscopy is highly operator-dependent. LCI enhances the visibility of colorectal lesions compared to WLI, regardless of the endoscopist’s level of experience.27,28
In a preliminary study of our group, ADR was significantly higher using LCI than WLI (56.9% and 43.2%, p=0.03), with LCI providing a marginal significance for SDR (p=0.05).29 Conceptually, ADR involved screening patients, although symptomatic and surveillance patients also participated in this study. Therefore, we published another study that enrolled only screened patients. The ADR was significantly higher in the LCI group than in the WLI group (71.0% and 52.9%, p=0.04), with no difference in SDR or AADR.12 The withdrawal time did not differ between the groups in either study.
Suzuki et al.,13 in a study of 3050 patients, reported a significantly higher ADR in the LCI group than in the WLI group (58.7% vs. 46.7%, p<0.01), as well as SDR (4.8% vs. 2.8%, p<0.01) and MAP (1.48 vs. 1.02, p<0.01), with no significant difference in AADR (13.2% vs. 11.0%, p=0.06). A recent meta-analysis of 17 RCTs and 10,624 patients showed significantly superior results with LCI compared with WLI for ADR, SDR, AADR, MAP, adenoma miss rate (AMR), and nonpolypoid lesions. The favorable effect of LCI on the detection of SSLs, AAs, and nonpolypoid lesions was observed only in studies which experts and trainees were involved but not for experts only.30 However, the meta-analysis conducted by Wang et al.31 demonstrated a significant increase in ADR (51.3% vs. 43.8%, p=0.0001) and reduction in AMR (12.2% vs. 24.4%, p=0.004) using LCI compared with WLI, with no difference in SDR or AADR.
The present study aimed to compare the performance of LCI colonoscopy and AI-assisted LCI colonoscopy and evaluate their impact on the detection of neoplasms. The primary application of AI in colonoscopy is detection (CADe), which improves the performance of inexperienced endoscopists and achieves results comparable to those obtained by experts. In fact, according to the European Society of Gastrointestinal Endoscopy, for acceptance of AI in the assessment of completeness of mucosal visualization and in the detection of colorectal polyps, the AI-assisted ADR should be comparable to that of experienced endoscopists.32
Miyaguchi et al.22 were the first to compare AI-assisted LCI with LCI alone and demonstrated significantly higher ADR (58.8% vs. 43.5%, p<0.001) and MAP (1.31 vs. 0.94, p<0.001) for AI-assisted LCI, especially in the ascending colon. The ADR for experts was 56.2% in the LCI+AI group and 46.2% in the LCI group (p=0.02), whereas for trainees, the ADR were 63.4% and 38.9% (p<0.0001) in the LCI+AI and LCI groups, respectively. The difference in ADR between experts and trainees was not significant in either group. The SDR was higher in the LCI+AI group (4.0% vs. 1.0%, p=0.007). Nonpolypoid adenomas were also more commonly detected in the LCI+AI group, as were small polyps (≤5 mm [p<0.001], 6–9 mm [p=0.04]).
By contrast, our study showed equivalent results for the detection of neoplasms when comparing LCI alone with AI-assisted LCI. The ADR was 48.0% vs. 52.6% (p=0.13), the SDR was 8.2% vs. 8.5% (p=1.0), and the NDR was 52.3% vs. 56.6% (p=0.30), respectively. The MAP was 0.9 for both groups. The AADRs were 5.9% for LCI alone and 5.7% for AI-assisted LCI. The main limitation of this study and its generalizability is that colonoscopies were performed by a single experienced endoscopist, as AI systems are theoretically promoted as tools for non-expert endoscopists. The high ADR in the LCl group may have masked potential benefits of AI by causing a statistical ceiling effect (48% vs. 52.6% ADR). However, significantly more adenomas were detected in the right colon in the LCI+AI group (63.8% vs. 52.1%, p=0.01). This may be explained by the fact that nonpolypoid lesions were detected more often in the right colon than in the left colon, and more nonpolypoid adenomas were detected in the LCI+AI group than in the LCI only group (70.3% vs.62.5%, p=0.06).
When screening, surveillance, and symptomatic patients were analyzed separately, the ADRs were similar between the LCI and LCI+AI groups. The smaller number of adenomas diagnosed in the symptomatic patients group is explained by the significantly larger number of patients aged <50 years in the symptomatic group than in the other groups.
Hassan et al.15 analyzed 21 RCTs involving 18,232 patients and reported a higher ADR in the CADe group than in the control group (44.0% vs. 35.9%), corresponding to a 55% relative reduction in AMR. In a meta-analysis of 24 RCTs involving 17,413 colonoscopies, Lee et al.16 showed that AI-assisted colonoscopy significantly increased the ADR, and this improvement was significantly higher in studies conducted in Asia than in those conducted in Europe or North America. No differences were observed between the AI systems. Studies including only expert endoscopists and those including both expert and nonexpert endoscopists showed comparable results, both reporting an increase in ADR.16 The meta-analysis conducted by Jin et al.33 showed a substantially lower miss rate for adenomas and SSLs in the AI group than in the non-AI group, but with no reduction in the AA miss rate. Biscaglia et al.34 compared the performance of AI-assisted trainees versus experts and obtained comparable results between both groups for ADR (38% vs. 40%) and MAP (0.93 vs. 1.07), with no significant differences in either parameter.
Failure to expose the mucosal folds is considered a cause of missed lesions, with a consequent relationship with interval CRC, which may be explained by rapid withdrawal of the colonoscope (<6 min), insufficient training, poor withdrawal technique, or even endoscopist fatigue. In a comparative analysis among AI, a single observer, and a dual observer, both AI and dual observers showed a higher ADR than a single observer, with no statistical difference between AI and the second observer.35 The high performance of CADe suggests that AI can act as a “second eye” for the endoscopist, reducing AMR.
The main limitation of this study is that all colonoscopies were performed in a single endoscopy center by the same endoscopist with expertise in IEE and some experience in AI. Further studies are required to evaluate whether AI can assist IEE in detecting neoplastic lesions.
In conclusion, this is the first RCT conducted in a western country to compare LCI and AI-assisted LCI colonoscopy, showing similar results in both groups, without detracting from the performance of AI-assisted LCI but evaluating high ADR using LCI alone, as well as good results in terms of SDR and NDR. Most adenomas had a nonpolypoid morphology in both groups, and significantly more adenomas were detected in the right colon in the LCI+AI group than in the LCI alone group.