JMIR Publications

Background:

Accurate real-time classification of colorectal polyps during colonoscopy is essential for risk stratification, therapeutic decision-making, and surveillance planning, with important implications for colorectal cancer prevention, health care costs, and patient burden. Despite advances in optical diagnosis techniques, including virtual chromoendoscopy and magnification, widespread adoption remains limited because of high operator dependency and variable diagnostic accuracy. Existing computer-assisted diagnosis (CADx) systems predominantly provide binary classification of colorectal lesions, thereby constraining their clinical utility for nuanced endoscopic decision-making.

Objective:

This study aims to develop and validate a real-time, multiclass AI-based CADx system capable of classifying colorectal lesions into 5 categories: nonneoplastic lesions, low-grade adenomas, high-grade adenomas, sessile serrated lesions (SSL), and invasive colorectal cancer.

Methods:

This multicenter international study consists of 2 phases: model development and proof-of-concept validation. During the model development phase, a large, diverse dataset of >17,000 deidentified endoscopic images and video frames will be prospectively collected from 8 European centers and 1 Japanese center, capturing a broad spectrum of lesion types and patient demographics. A semisupervised learning framework using contrastive representation learning (SimCLR [simple framework for contrastive learning of visual representations]) will be used to leverage both labeled and unlabeled data, enhancing feature extraction and classification performance. Expert consensus annotation will be used for labeled data. Lesion localization will be supported by an integrated computer-assisted detection (CADe) module, while model interpretability will be facilitated through gradient-weighted class activation mapping (Grad-CAM). Proof-of-concept validation will be conducted using 250 previously unseen lesion images that will be independently assessed by expert endoscopists. The primary outcome is multiclass diagnostic accuracy, with a predefined target accuracy threshold of 85%, alongside sensitivity and specificity for invasive cancer detection. Secondary outcomes include interobserver agreement, performance across demographic and lesion subgroups, and a qualitative analysis of misclassification patterns. All data will be securely managed within a REDCap database, with strict deidentification protocols and plans to generate an open-access research resource.

Results:

Ethics approval has been obtained at all participating centres. Data transfer agreements have been executed by the majority of collaborating sites, with de-identified polyp images and videos currently being uploaded to a secure central database.

Conclusions:

This study aims to deliver a clinically applicable, real-time multiclass CADx system that extends beyond current binary AI models and supports more granular, personalized endoscopic decision-making. By integrating semisupervised learning, diverse international data, and explainable AI techniques, this approach has the potential to substantially advance the methodological foundations and clinical utility of AI-assisted optical diagnosis in routine colonoscopy, contributing to improved colorectal cancer prevention and patient care.