Narrow Band Imaging Versus Artificial Intelligence for Colonic Surveillance in Inflammatory Bowel Disease: a Prospective, Randomized, Crossover Study The CLEAR-IBD Trial (Colorectal Lesion Evaluation Assisted by Real-time AI in Inflammatory Bowel Disease)
Narrow Band Imaging Versus Artificial Intelligence for Colonic Surveillance in Inflammatory Bowel Disease: a Prospective, Randomized, Crossover Study The CLEAR-IBD Trial (Colorectal Lesion Evaluation Assisted by Real-time AI in Inflammatory Bowel Disease)
Patients with inflammatory bowel disease (IBD) - ulcerative colitis or Crohn's disease - have a higher risk of developing colorectal cancer than the general population. For this reason, regular colonoscopies are recommended to look for early warning signs, such as abnormal areas of tissue called dysplasia, which can develop into cancer over time.
Finding these abnormal areas during colonoscopy can be difficult because IBD causes ongoing inflammation that can make the bowel mucosa look irregular, hiding subtle changes. Doctors currently use a technique called narrow-band imaging (NBI), a special light setting on the colonoscope that enhances the visibility of the bowel, to help spot these areas more easily.
A newer tool, artificial intelligence (AI)-assisted detection, has shown promise in helping doctors find more polyps during routine colonoscopies in the general population. However, this AI tool was developed and tested mostly in people without IBD, so it is not yet known whether it works as well in people with IBD, whose bowel can look very different due to chronic inflammation.
This study will directly compare the AI tool (CADe; ENDO-AID, Olympus) with narrow-band imaging to see which one is better at finding abnormal areas during colonoscopy in patients with long-standing ulcerative colitis or Crohn's disease who are having their routine cancer surveillance exam.
Each participant will have one colonoscopy in which the bowel is examined twice in a row, once with each technique, by two different doctors, in random order. This lets researchers compare both methods directly within the same patient, which is the fairest comparison.
The study aims to enroll 60 patients at Vall d'Hebron University Hospital in Barcelona, Spain. Researchers hope the results will help determine whether AI tools - which are widely available and easier to use than NBI - can be a reliable alternative for IBD surveillance, potentially making this important cancer-screening exam more accessible in the future.
Background and Rationale
Patients with inflammatory bowel disease (IBD) remain at increased risk of colorectal cancer (CRC) compared with the general population, and colonoscopic surveillance has been shown to reduce both CRC incidence and CRC-related mortality. Although high-definition dye-based chromoendoscopy is recommended by most guidelines as the surveillance technique of choice, its uptake remains limited due to training requirements, added procedure time, and cost. As a result, several guidelines continue to accept virtual chromoendoscopy techniques, including narrow-band imaging (NBI), and high-definition white-light endoscopy as alternatives.
Computer-aided detection (CADe) systems based on artificial intelligence have demonstrated meaningful improvements in adenoma and polyp detection in non-IBD populations. However, the convolutional neural network-based algorithms underlying these systems have largely been trained on datasets that excluded patients with IBD, so the real-world performance of currently available non-IBD-trained CADe systems for detecting colitis-associated dysplasia remains unknown. Existing evidence in this area has relied mainly on retrospective analyses of still-image datasets, without prospective, real-time comparison against validated comparators such as chromoendoscopy or NBI. A direct, prospective comparison is therefore needed.
Study Design This is a unicentric, two-arm, open-label, randomized, crossover clinical trial conducted at the IBD Clinic of the Gastroenterology Department, Vall d'Hebron University Hospital, Barcelona, Spain. Each participant undergoes a single colonoscopy in which the entire colon is examined twice, consecutively, by two different endoscopists using the two study techniques (CADe and NBI) in randomized order, allowing each patient to act as their own control and minimizing between-patient confounding.
Study Procedures Each participant is examined on the same day by two endoscopists with expertise in IBD endoscopic surveillance. Endoscopist 1 examines the entire colon using either AI (ENDO-AID CADe System, Olympus, Tokyo, Japan) or NBI, per the randomization sequence. Before lesion assessment, endoscopic remission is confirmed using the Mayo endoscopic score (UC) or the Simple Endoscopic Score for Crohn's Disease (SES-CD), requiring a score of 0. Lesions identified are documented and managed accordingly (resection or biopsy). Endoscopist 2, blinded to the findings of Endoscopist 1, re-examines the entire colon using the alternate technique, documents any additional findings, and manages any lesions missed during the first examination. Both endoscopists perform an equal number of examinations with each technique as the first or second pass, and all procedures use the same high-definition endoscope and processor (Olympus CF-HQ190L, EVIS X1).
All lesions resected or biopsied, regardless of detection order, are categorized using the Paris classification, with location and size recorded. Specimens are assessed by pathologists with expertise in IBD as part of standard care; adenocarcinoma, serrated adenoma, tubular adenoma, and any dysplasia are classified as neoplastic lesions. Biopsies are systematically obtained from previously identified colitic segments in all patients to assess the impact of histological inflammatory activity on neoplasia detection. Withdrawal time, defined as the interval between the start of inspection at the cecum and the end of the procedure, is recorded separately for each technique and examination order.
Randomization and Blinding Patients are consecutively allocated in a 1:1 alternating sequence to either the "NBI first followed by AI" arm or the "AI first followed by NBI" arm, with allocation performed by Endoscopist 1 prior to the procedure. Blinding of endoscopists to the technique itself is not feasible given that the AI overlay is visible on screen; however, Endoscopist 2 remains blinded to the findings of the first examination, preserving the integrity of the tandem comparison.
Data Collection Patient and disease characteristics collected include age, sex, IBD type, age at diagnosis, disease location and behavior (Montreal classification for CD; extent for UC), presence of perianal disease, presence of primary sclerosing cholangitis, current and past medication, personal history of dysplasia, previous IBD-related surgery, and time elapsed since the last surveillance colonoscopy. Procedural data collected include the randomization arm, order of techniques, all detected lesions (technique, endoscopist, morphology, size, location), histological findings, withdrawal times, and any procedure-related complications.
Sample Size and Statistical Analysis In the absence of prior data on CADe performance in IBD surveillance colonoscopy, the sample size was pragmatically set at 60 evaluable patients, based on previously published crossover studies comparing dye-based chromoendoscopy and NBI in this population. A 20% dropout rate was anticipated to account for incomplete explorations, withdrawal of consent, and suboptimal bowel preparation, so recruitment was planned for approximately 72 patients. Lesion-level detection (paired binary outcomes) will be compared using the McNemar test, while per-patient dysplastic lesion counts will be compared using the Wilcoxon signed-rank test. Because most lesions are sampled immediately upon detection, only missed or unresected lesions can be identified during the second examination, precluding fully paired analyses for some comparisons; individual lesions are therefore treated as statistically independent observations in those cases. Quantitative variables are compared using Student's t-test or the Mann-Whitney U test as appropriate, and proportions using the χ² test or Fisher's exact test, with Bonferroni correction applied for multiple comparisons involving lesion location and histology. All p-values are two-sided, with p < 0.05 considered statistically significant. Analyses are performed using SPSS version 16 (SPSS Inc., Chicago, IL).
Ethical and Regulatory Considerations This protocol was approved by the Drug Research Ethics Committee (Comité de Ética de la Investigación con Medicamentos) of Hospital Universitari Vall d'Hebron under code PR(AG)237/2023. All participants provide written informed consent prior to enrolment, and the study is conducted in accordance with the Declaration of Helsinki and applicable national and institutional regulations. No patients or members of the public were involved in the design, conduct, reporting, or dissemination of this research. Any immediate or delayed procedure-related complications are recorded and reported according to institutional adverse event monitoring protocols.
Funding This work is supported by the Societat Catalana de Digestologia through the Ajut per a la Iniciació a la Recerca grant. The funder has no role in study design, data collection, analysis, interpretation, or manuscript decisions. The investigators declare no relevant conflicts of interest.
Inclusion Criteria:
Exclusion Criteria: