AI-Enhanced Prognostic Model for Predicting Polyp Recurrence and Guiding Post-Polypectomy Surveillance Intervals...

Area of Science:

• Gastroenterology and AI
• Medical Imaging Analysis
• Oncology

Background:

• Colorectal cancer (CRC) is a significant health concern, with adenomatous polyps as precursors.
• Post-polypectomy recurrence poses a substantial risk for CRC development.
• Current surveillance strategies require optimization for personalized patient management.

Purpose of the Study:

• To develop and evaluate a multimodal AI model for predicting colorectal polyp recurrence post-polypectomy.
• To integrate endoscopic imaging with clinical and pathology data for enhanced risk prediction.
• To support individualized colonoscopy surveillance planning based on recurrence risk.

Main Methods:

• A multimodal AI model was developed using the ERCPMP-v5 dataset, incorporating endoscopic images and structured clinical/pathology data.
• A late fusion framework combined features from a Vision Transformer (ViT) backbone and a multilayer perceptron for metadata.
• Model performance was assessed using accuracy, precision, recall, F1-score, and AUC, compared against image-only and metadata-only models.

Main Results:

• The multimodal fusion model achieved high performance: 90.4% accuracy, 86.7% precision, 83.1% recall, 84.9% F1-score, and 0.920 AUC.
• Multimodal fusion outperformed image-only (0.880 AUC) and metadata-only (0.850 AUC) models.
• Risk stratification generated tailored surveillance intervals: 1-3 years (low), 6-12 months (moderate), and 3-6 months (high risk).

Conclusions:

• A multimodal AI model integrating endoscopic imaging and clinical data effectively predicts post-polypectomy recurrence.
• The model provides actionable, risk-based surveillance recommendations for individualized patient follow-up.
• This approach enhances surveillance efficiency and resource allocation while prioritizing high-risk patients.