Document resource
Background Flat polyps cause the majority of interval colorectal cancers yet remain the most AI-missed lesion subtype. We quantified the flat polyp detection gap in a state-of-the-art AI system, evaluated focal loss reweighting as a mitigation strategy, and modelled data requirements for clinically acceptable performance.Methods Polyps from six PolypGen centres (n=2,020 frames) were classified as flat (area <1% of frame; n=344, 17.0%) or protruding (n=1,676, 83.0%) ( IDDF2026-ABS-0261 Figure 4(a) Flat vs protruding polyp characteristics in clinical risk context). RGB contrast between polyp and surrounding mucosa was measured as a visual detectability proxy (IDDF2026-ABS-0261 Figure 3. Visual colour cue analysis comparing flat and protruding polyps). A YOLOv8 baseline was compared against a focal-loss reweighted model (cls=2.0) with 3× flat polyp oversampling, evaluated on two held-out centres (C3-UK, C4-Egypt).Results Flat polyps demonstrated 28% lower mucosal contrast than protruding lesions (μ=116.2 vs 161.7 RGB units), with smaller frame area and more circular morphology confirming their inherent visual difficulty (IDDF2026-ABS-0261 Figure 3. Visual colour cue analysis comparing flat and protruding polyps, IDDF2026-ABS-0261 Figure 4. Flat vs protruding polyp characteristics in clinical risk context). Training data composition revealed only 344 flat frames versus 1,676 protruding a fundamental imbalance driving persistent detection bias (IDDF2026-ABS-0261 Figure 1(a) Three-panel analysis of flat polyp detection barriers and data requirements).The baseline model detected 50.3% of flat polyps versus 87.6% of protruding lesions (gap=37.3pp; χ2=101.92, p<0.001). Focal reweighting with oversampling failed to improve flat recall (48.4% vs 50.3%; Δ= -1.9pp) while modestly improving protruding detection (89.8%; Δ=+2.2pp), with flat detection rates consistent across both held-out centres (IDDF2026-ABS-0261 Figure 2. Focal loss reweighting results across three panels). Learning curve modelling estimated ≥1,720 flat polyp training frames 5× current availability are required to reach the 80% clinical detection threshold (IDDF2026-ABS-0261 Figure 1(c) Three-panel analysis of flat polyp detection barriers and data requirements). Even with 5× more flat data, modelling projects flat recall reaching only 80% versus 90% for protruding lesions (IDDF2026-ABS-0261 Figure 1(b) Three-panel analysis of flat polyp detection barriers and data requirements).Conclusions Flat polyp detection is fundamentally limited by data poverty rather than algorithm design. With only 17% flat polyp representation, focal loss reweighting is insufficient to bridge the detection gap. Achieving clinically acceptable flat polyp detection requires dedicated large-scale flat polyp data collection, the most urgent unmet need in colonoscopy AI development.Abstract IDDF2026-ABS-0261 Figure 1Abstract IDDF2026-ABS-0261 Figure 2Abstract IDDF2026-ABS-0261 Figure 3Abstract IDDF2026-ABS-0261 Figure 4