JMIR Publications

ABSTRACT

Accurate classification of colorectal polyps during colonoscopy is critical for determining individual cancer risk, guiding treatment, and optimising surveillance, thereby reducing healthcare costs and patient burden. Despite advances in optical diagnosis—such as virtual chromoendoscopy and magnification—its adoption remains limited due to skill requirements and suboptimal accuracy. Current computer-assisted diagnosis (CADx) tools primarily offering binary (neoplastic vs. non-neoplastic) classification of colorectal polyps. To address these limitations, we propose developing a multiclass AI-based CADx system capable of real-time, five-class classification of colorectal lesions, including non-neoplastic lesions, low- and high-grade adenomas, sessile serrated lesions (SSLs), and invasive cancers. This multicentre, international study comprises two phases: development and validation. During the development phase, we will collect and curate a comprehensive dataset of over 17,000 de-identified endoscopic images and videos, sourced prospectively from eight European hospitals and one centre in Japan. The dataset encompasses diverse ethnicities and lesion types, facilitating robust AI training. A semi-supervised learning approach utilising contrastive learning (SimCLR) will be employed to optimise the model’s discriminative capability. Labelling of images will involve consensus from experienced endoscopists, while unlabelled data will be leveraged through contrastive pre-training, enhancing feature extraction even with limited annotations. An integrated CADe component will detect lesions, with heatmaps providing interpretability via Grad-CAM to highlight relevant image features. The proof-of-concept validation will involve 250 unseen lesion images scored by expert endoscopists, with a target accuracy threshold of 85%. Primary outcomes include sensitivity and specificity for cancer prediction, alongside secondary measures such as interobserver agreement, diagnostic accuracy across demographic subgroups, and qualitative analyses of misclassifications. Data management will adhere to strict de-identification protocols, housed within a secure REDCap database, with an open-access resource generated for future research. Ethics approval has been secured or is pending across participating sites, with informed consent obtained prospectively and opt-out options available retrospectively. The proposed outcome of this project is a real-time AI tool that provides detailed lesion classification, surpassing current binary systems and supporting personalised endoscopic decision-making. This pioneering effort promises to advance CADx utility in routine practice, ultimately contributing to improved colorectal cancer prevention and patient care worldwide.