Accurate detection of submucosal invasive cancer (SMIC) in large non-pedunculated colorectal polyps (LNPCPs) is essential for risk stratification. Pre-resection videos are increasingly used for expert review and triage, but it is unknown whether video quality influences diagnostic performance. This prospective multicentre study evaluated the impact of video quality on SMIC detection in LNPCPs by trained endoscopists.

In four tertiary centres, high-definition pre-resection videos of consecutive LNPCPs were recorded by expert endoscopists and optically trained interventional fellows. Videos were uploaded to a web platform and independently reviewed by 9 blinded raters (experts and fellows) not involved in the original procedures. Raters were randomly assigned to videos and classified each lesion for SMIC presence.

Video quality was scored on a 5-point Likert scale (1=poor, 5=excellent) following a standardisation meeting defining structural criteria (global white-light overview, systematic inspection with virtual chromoendoscopy, close-up of suspicious areas) and technical criteria (clean lens, adequate washing, steady images). All recording endoscopists had received a brief educational intervention on standardised LNPCP video acquisition to ensure a basic proficiency in pre-resection imaging. 

The primary outcome was the association between video quality and diagnostic performance (sensitivity, specificity, accuracy) for SMIC detection as defined by blinded histopathological assessment. Video quality was analysed as ordinal (1–5) and categorically as low (1–3) versus high (4–5). Sensitivity, specificity and accuracy were calculated with 95% confidence intervals (CI). A mixed-effects logistic regression model with correct SMIC classification as dependent variable and random intercepts for lesion and rater assessed quality impact.

We included 268 LNPCPs, yielding 667 video assessments. SMIC was present in 53/268 (19.8%), HGD in 76/268 (28.4%), and LGD in 129/268 (48.1%). Videos were rated as low-quality in 184/667 (27.6%) and high-quality in 483/667 (72.4%).  

Overall sensitivity, specificity and accuracy for SMIC detection were 70.0% (95%CI 61.7–77.4), 83.8% (95%CI 80.4–86.7) and 81.0% (95%CI 77.9–83.8), respectively. Stratified by quality category, sensitivity was 67.0% (95%CI 56.7–76.2) for high-quality versus 73.0% (95%CI 55.9–86.2) for low-quality videos (P=0.648). Specificity was 84.5% (95%CI 80.4–87.9) vs 81.0% (95%CI 73.7–87.0) (P=0.399), and accuracy 81.0% (95%CI 77.2–84.4) vs 79.3% (95%CI 72.8–85.0) (P=0.720). Lesion demographics are presented for high- and low-quality videos in the table.  In the mixed-effects model, each 1-point increase in quality score yielded OR 1.07 (95%CI 0.70–1.65; P=0.76) for correct classification, indicating no significant effect.

In this multicentre study, SMIC detection on pre-resection LNPCP videos was highly accurate and did not differ significantly between low- and high-quality videos when basic structural and technical criteria were met. These findings support the use of routine pre-resection videos of varying quality for risk stratification and MDT triage by trained endoscopists.