Document resource
Background Pathological assessment of all diminutive colorectal polyps (≤5mm) is resource-intensive and unnecessary for hyperplastic lesions. AI-based ‘resect and discard’ strategies could reduce pathology burden, but existing approaches rely on optical chromoendoscopy features unavailable in standard white-light endoscopy. Morphological shape analysis offers a modality-agnostic alternative applicable to any segmentation-capable colonoscopy system. We developed an explainable AI model predicting diminutive polyp histology from morphological shape features derived from segmentation masks, and evaluated whether performance meets ASGE PIVI criteria (NPV ≥90%).Methods 1,000 polyp images from Kvasir-SEG with expert segmentation masks were analysed. Seventeen shape descriptors were extracted per polyp including circularity, fractal dimension, solidity, and convex hull defects. Four classifiers (XGBoost, Random Forest, Logistic Regression, SVM) were trained on a 70/10/20 train/val/test split with isotonic probability calibration applied post-hoc. SHAP values provided per-prediction explainability at the global and case level.Results t-SNE visualisation of the shape feature space demonstrated meaningful separation between hyperplastic and adenomatous polyps, with overlap confined to morphologically ambiguous cases ( IDDF2026-ABS-0262 Figure 1. t-SNE visualisation of the 17-variable shape feature space with KDE contour overlays for hyperplastic and adenomatous polyps). XGBoost achieved AUC=0.998, sensitivity=0.973, specificity=0.966, NPV=0.966, and Expected Calibration Error=0.043, exceeding the ASGE PIVI threshold of NPV≥0.90. Random Forest also met PIVI criteria (NPV=0.955, AUC=0.995), while Logistic Regression (NPV=0.859) and SVM (NPV=0.875) did not (IDDF2026-ABS-0262 Figure 3. ROC curves for all four classifiers, IDDF2026-ABS-0262 Figure 4. NPV vs AUC comparison across all four models against the PIVI threshold). SHAP analysis identified polyp size (mean |SHAP|=2.61) and surface complexity reflecting pit pattern (mean |SHAP|=1.62) as dominant predictors, with shape regularity, convex defect count, and lobulation as secondary contributors (IDDF2026-ABS-0262 Figure 2. Global shape feature importance for the XGBoost model). Case-level SHAP waterfall plots demonstrated individual prediction explainability suitable for clinical adoption.Conclusions Morphological shape features derived from polyp segmentation masks predict histology with PIVI-compliant accuracy using an explainable AI framework. XGBoost and Random Forest both exceed the NPV≥0.90 threshold, supporting resect-and-discard implementation without requiring optical chromoendoscopy, with direct potential for integration into existing AI-assisted colonoscopy pipelines. External validation in real-world colonoscopy cohorts is warranted before clinical deployment.Abstract IDDF2026-ABS-0262 Figure 1Abstract IDDF2026-ABS-0262 Figure 2Abstract IDDF2026-ABS-0262 Figure 3Abstract IDDF2026-ABS-0262 Figure 4