Patients with IBD are at increased risk for dysplasia and colorectal cancer (CRC) often at a younger age than the average population, making early detection essential. Despite routine surveillance colonoscopy, post-colonoscopy CRC rates remain high (28–41%), partly due to the subtle morphology and rapid progression of IBD-associated lesions, as well as the technical challenges in surveillance colonoscopy. Therefore, simpler and more effective endoscopic techniques are needed to improve dysplasia detection. The role of artificial intelligence (AI) in lesion detection during colonoscopy is well established in patients without IBD, but remains unclear in those with IBD. 

A systematic review has been carried out to assess the diagnostic yield of AI software for detecting dysplasia in patients with IBD, specifically CADe systems designed for dysplasia detection.

A Systematic literature searches was conducted on Medline, Cochrane Library, Web of Science, IEEE, ACM, and Scopus between 2015-2025. All retrieved references were imported to the Rayyan software for assessment for eligibility. The quality of included studies was assessed using the QUADAS-2 tool. 

The systematic search identified 3032 studies, which were screened through multiple stages. Duplicates (n =1408) were identified and removed, and a further 1,618 studies were excluded for not meeting the inclusion criteria (e.g., not addressing the study question or not being original research). Ultimately, three studies specifically investigated the detection of dysplasia during surveillance colonoscopy in patients with IBD. However, due to substantial heterogeneity in study design— including prospective real-time lesion detection in colonoscopy videos, image-based lesion detection, lesion classification, and image-based AI model development —the results could not be pooled for a meta-analysis. Therefore, a descriptive analysis was performed instead. An additional three studies explored related topics, including dysplasia classification, polyp detection, and comparisons among different AI filters. 

In two studies (Abdelrahim et al. 2024 and Vinsard et al 2023), researchers developed IBD-specific CADe systems by retraining existing CADe models using lesions derived from patients with IBD. Abdelrahim et al validated an IBD-dedicated CADe system using both image-based and real-time validation. In image-based validation, the IBD-specific model demonstrated higher performance compared with a generic CADe system. During real-time evaluation, the system achieved a lesion detection rate of approximately 90.4%.  Vinsard et al validated an IBD-CADe model exclusively using images of IBD-associated lesions. This system, developed with HDWLE images, demonstrated a higher sensitivity, specificity, PPV, and NPV than the original CADe model. 

Another study (López-Serrano, A. et al 2024) performed a cross-sectional non-inferiority comparison of Virtual chromoendoscopy (VCE) iScan with CADe Discovery in ulcerative colitis using histopathology as the gold standard. Among 52 eligible patients, 61 lesions were detected. CADe Discovery demonstrated a lesion detection rate similar to VCE iScan.

Additional studies evaluated multiple AI-based filters to determine the most effective filter to identified IBD-related lesions and polyps, contributing to optimization of AI feature selection for IBD surveillance.

A pilot study assessed the capability of a CADe system to differentiate low-grade from high-grade dysplasia and compared it both expert and non-expert endoscopists performance, providing early evidence for AI-assisted dysplasia grading.

Few studies have explored the diagnostic yield of AI for the detection and characterization of dysplastic lesions in IBD surveillance colonoscopy.  The lack of standardized protocols and heterogenicity, make comparisons between studies and pooled analyses impossible. Thus, more uniformed study designs are required.