{
  "abstract": "Background Incomplete polypectomy contributes to approximately 20% of interval colorectal cancers. Current AI systems detect and segment polyps but provide no quantitative ­assessment of resection margin confidence, leaving endoscopists without guidance on whether excision boundaries are adequate.Methods A DeepLabV3+ architecture (EfficientNet-B4 encoder, ImageNet pre-training) was trained on Kvasir-SEG (n=1,000 expert-annotated polyp images). Monte Carlo Dropout (p=0.3, n=25 forward passes) at the encoder bottleneck generated calibrated per-pixel uncertainty estimates. A boundary-aware loss function applied 3× penalty weighting at polyp margins. Margin zones were automatically classified as safe (σ≤0.07), caution (0.07<σ≤0.15), or danger (σ>0.15) based on MC Dropout standard deviation. Primary outcomes were Dice Similarity Coefficient (DSC) and uncertainty calibration assessed by Pearson correlation between uncertainty and boundary prediction error.Results Training converged stably over 50 epochs with no overfitting, achieving the best validation Dice of 0.9106 ( IDDF2026-ABS-0263 Figure 2. Polypectomy margin model training curves across 50 epochs. Left-Training and validation loss demonstrating stable convergence Right- Validation Dice score trajectory reaching best performance). Uncertainty estimates were well-calibrated, with model uncertainty strongly correlated with boundary prediction error (Pearson r=0.786, p<0.001; IDDF2026-ABS-0263 Figure 1. Uncertainty calibration plot showing Pearson correlation), confirming that high-uncertainty zones correspond to genuinely ambiguous margins. Three-zone confidence overlays, safe (green), caution (yellow), and danger (red), were generated in real time (<40ms/frame), with uncertainty concentrating precisely at polyp-mucosa boundaries across diverse morphologies (IDDF2026-ABS-0263 Figure 3. Uncertainty-aware AI margin confidence maps for three representative cases across the uncertainty spectrum). Clinically, Case A demonstrated safe resection margins (safe margin 92%, danger 0%), Case B showed moderate boundary uncertainty warranting caution (safe margin 89%), and Case C flagged high-risk margins requiring additional resection consideration (safe margin 87%, danger 1%) (IDDF2026-ABS-0263 Figure 3. Uncertainty-aware AI margin confidence maps for three representative cases across the uncertainty spectrum).Conclusions Uncertainty-aware AI margin confidence maps are technically feasible, computationally efficient, and well-calibrated. These maps provide endoscopists with interpretable, real-time resection guidance beyond binary polyp detection. Prospective validation against histopathological margin status (R0/R1) is warranted to establish clinical utility in reducing incomplete polypectomy rates.Abstract IDDF2026-ABS-0263 Figure 1Abstract IDDF2026-ABS-0263 Figure 2Abstract IDDF2026-ABS-0263 Figure 3",
  "authors": [
    {
      "affiliations": [
        "New York Medical College, United States"
      ],
      "name": "Jeril Lasington"
    },
    {
      "affiliations": [
        "Rutgers University, United States"
      ],
      "name": "Lawin Steve Mathew Lasington"
    },
    {
      "affiliations": [
        "Boston University, United States."
      ],
      "name": "Swamynathan Umamaheshwaran"
    }
  ],
  "title": "IDDF2026-ABS-0263 The margin the eye cannot see: AI uncertainty mapping for colonoscopic polypectomy",
  "uid": "f6dfbc1a-5b6e-5715-bd98-3e1579758c67"
}