Accurate real-time optical characterisation of colorectal polyps during colonoscopy is essential for optimal management and colorectal cancer prevention. Reliable differentiation of histological subtypes guides treatment decisions, determines surveillance intervals, and reduces unnecessary surgical resections. Conventional optical diagnosis using chromoendoscopy or virtual imaging is limited by operator dependence, and steep learning. Current computer-aided diagnosis (CADx) systems assist endoscopists but currently limited to binary classification (neoplastic vs non-neoplastic).  

Our aim is to develop a CADx algorithm capable of differentiating polyps into four clinically relevant categories with different options for treatment decisions and follow-up: hyperplastic polyps, sessile serrated lesions (SSLs), adenomas, and T1 cancers, 

Between May 2023 to June 2025, we assembled an image dataset comprising of 8,635 unique polyps in white light image (WLI) and / or Narrow band image (NBI) from 6 institutions across five countries (United Kingdom, Spain, Germany, Norway, and Japan). The dataset included 1,455 hyperplastic polyps, 707 SSLs, 5,767 adenomas (low- and high-grade combined), and 706 T1 cancers. A two-stage classification framework using the ConvNext-Small convolutional neural network architecture was implemented. In stage one, a binary classifier distinguished serrated-type (hyperplastic and SSL) from adenomatous-type (adenoma and T1 cancer) lesions. In stage two, two separate binary models were trained: a serrated model (hyperplastic vs SSL) and a depth model (adenoma vs T1 cancer). Model performance was assessed using an independent test set of 231 images (11 hyperplastic, 41 SSL, 121 adenoma, and 58 T1 cancer). 

The multiclass model achieved an overall four-class classification accuracy of 68.8%. When grouped into serrated versus adenomatous lesions, diagnostic accuracy improved to 84.4%, with a sensitivity of 93.9% and specificity of 52.0%. The depth model differentiating adenomas from T1 cancers achieved an accuracy of 79.8%, and specificity of 80.6%.

This study demonstrates the feasibility of a four-class CADx algorithm for endoscopic differentiation of hyperplastic polyps, SSLs, adenomas, and T1 cancers, extending beyond conventional binary AI systems. This approach has the potential to support real-time therapeutic decision-making and improve patient-level risk stratification. By enabling tailored management strategies and reducing reliance on histopathology, multi-class CADx may contribute to more efficient, cost-effective, and patient-centred colorectal cancer prevention. Further validation using video-based datasets and multicentre prospective evaluation is warranted before clinical implementation.

Funding: 

This study is funded by the European Commission - Horizon Europe 101057099.