Multicentre prospective evaluation of real-time optical diagnosis of T1 colorectal cancer in large non-pedunculated colorectal polyps using narrow band imaging (the OPTICAL study).

Objective This study evaluated the preresection accuracy of optical diagnosis of T1 colorectal cancer (CRC) in large non-pedunculated colorectal polyps (LNPCPs). Design In this multicentre prospective study, endoscopists predicted the histology during colonoscopy in consecutive patients with LNPCPs using a standardised procedure for optical assessment. The presence of morphological features assessed with white light, and vascular and surface pattern with narrow-band imaging (NBI) were recorded, together with the optical diagnosis, the confidence level of prediction and the recommended treatment. A risk score chart was developed and validated using a multivariable mixed effects binary logistic least absolute shrinkage and selection (LASSO) model. Results Among 343 LNPCPs, 47 cancers were found (36T1 CRCs and 11 >= T2 CRCs), of which 11 T1 CRCs were superficial invasive T1 CRCs (23.4% of all malignant polyps). Sensitivity and specificity for optical diagnosis of T1 CRC were 78.7% (95% CI 64.3 to 89.3) and 94.2% (95% CI 90.9 to 96.6), and 63.3% (95% CI 43.9 to 80.1) and 99.0% (95% CI 97.1 to 100.0) for optical diagnosis of endoscopically unresectable lesions (ie, >= T1 CRC with deep invasion), respectively. A LASSO-derived model using white light and NBI features discriminated T1 CRCs from non-invasive polyps with a cross-validation area under the curve (AUC) of 0.85 (95% CI 0.80 to 0.90). This model was validated in a temporal validation set of 100 LNPCPs (AUC of 0.81; 95% CI 0.66 to 0.96). Conclusion Our study provides insights in the preresection accuracy of optical diagnosis of T1 CRC. Sensitivity is still limited, so further studies will show how the risk score chart could be improved and finally used for clinical decision making with regard to the type of endoresection to be used and whether to proceed to surgery instead of endoscopy.