{
  "abstract": "Background AI colonoscopy systems trained on European cohorts may underperform in diverse populations, limiting global adoption. We quantified performance degradation across PolypGen’s six international centres and characterised the morphological and colorimetric features driving population-specific detection bias.Methods A YOLOv8 detection model was trained on four European centres (C1-Norway, C2-France, C5-Italy, C6-­Norway; n=1,306 polyp frames) and evaluated on a held-out non-­European centre (C4-Egypt; n=153 frames). Per-centre morphological analysis quantified polyp size distribution, ­flatness prevalence, and mucosal colour profiles across all six centres. Flat polyps were defined as occupying <1% of the frame area.Results Centre-wise dataset composition revealed substantial heterogeneity in frame counts, polyp size distributions, and positive/negative frame ratios across all six centres ( IDDF2026-ABS-0260 Figure 1. Polypgen centre-wise overview across C1-C6). The European-trained model achieved mAP@50 of 0.963-0.991 and recall of 0.923-0.970 across training centres, degrading substantially on the non-European held-out centre (C4-Egypt: mAP@50=0.702, recall=0.618, F1=0.694), a 29% absolute reduction in detection accuracy.Morphological analysis revealed significant inter-centre variation in flat polyp prevalence, ranging from 8.3% to 24.2% across centres (IDDF2026-ABS-0260 Figure 2. Polyp morphology by centre as bias indicators for the Asian Pacific detection gap). Critically, Asian-Pacific centres are entirely absent from PolypGen, despite Asian polyps being 70-80% flat, threefold exceeding the maximum flat polyp rate in any available centre (IDDF2026-ABS-0260 Figure 2. Polyp morphology by centre as bias indicators for the Asian Pacific detection gap). Mucosal colour analysis demonstrated substantial RGB variation across centres (mean redness: 116.8-153.9), reflecting differences in endoscope hardware, bowel preparation, and population mucosal characteristics (IDDF2026-ABS-0260 Figure 3. Mucosal colour analysis by centre across RGB channels), all sources of domain shift invisible to a European-trained model.Conclusions European-dominant training datasets introduce systematic detection bias against morphologically distinct polyp populations. Models trained on PolypGen are predicted to underperform in Asian-Pacific colonoscopy programmes where flat, right-sided lesions predominate, precisely the highest-risk lesions for interval cancer. Dataset diversification incorporating Asian-Pacific centres is urgently needed to ensure equitable AI-assisted colonoscopy globally.Abstract IDDF2026-ABS-0260 Figure 1Abstract IDDF2026-ABS-0260 Figure 2Abstract IDDF2026-ABS-0260 Figure 3",
  "authors": [
    {
      "affiliations": [
        "Rutgers University, United States"
      ],
      "name": "Lawin Steve Mathew Lasington"
    },
    {
      "affiliations": [
        "New York Medical College, United States"
      ],
      "name": "Jeril Lasington"
    },
    {
      "affiliations": [
        "Boston University, United States"
      ],
      "name": "Swamynathan Umamaheshwaran"
    }
  ],
  "title": "IDDF2026-ABS-0260 Why colonoscopy AI may fail Asian-pacific patients",
  "uid": "99515e03-03ad-5849-87b6-fb0aa7dc8250"
}