Related Experiment Video
Updated: Aug 5, 2026
07:22
Surgical Robot-Assisted Transanal Specimen Extraction Radical Sigmoidectomy Without an Auxiliary Abdominal Incision
Published on: June 13, 2025
PolypSAM-Open: Mitigating Automation Bias in AI-Assisted Colonoscopy via Open-Set Surgical Artifact Rejection
Umar Hasan1, Shadman Shahriar1, Faiyad Hossain1
1Department of Electrical and Computer Engineering, School of Engineering and Physical Sciences, North South University, Dhaka 1229, Bangladesh.
Diagnostics (Basel, Switzerland)
|July 28, 2026
Summary
New AI framework PolypSAM-Open enhances colonoscopy by accurately identifying polyps and rejecting surgical tools. This improves workflow reliability and patient safety in AI-assisted procedures.
Area of Science:
- Medical Artificial Intelligence
- Computer Vision in Endoscopy
- Surgical Instrument Recognition
Background:
- AI decision support systems in colonoscopy risk failure with out-of-distribution instruments, causing false polyp detections.
- This automation bias can decrease the reliability of AI-assisted colonoscopy workflows.
- Developing robust AI that distinguishes polyps from instruments is crucial for clinical adoption.
Purpose of the Study:
- To develop and evaluate a parameter-efficient framework for open-set-aware polyp segmentation.
- The framework aims to reject anomalous surgical instrument inputs while maintaining accurate polyp segmentation.
- Enhance the reliability and safety of AI-assisted colonoscopy.
Main Methods:
- Proposed PolypSAM-Open, integrating Open-Set Learning (OSL) with Low-Rank Adaptation (LoRA) in the MedSAM encoder.
- Trained on Kvasir-SEG with polyp images and synthetic noise to establish a rejection margin.
- Evaluated zero-shot out-of-distribution detection on Kvasir-Instrument and compared segmentation performance against a baseline.
Main Results:
- PolypSAM-Open significantly improved zero-shot OOD detection AUROC from 0.4263 to 0.9535 (p<0.001) for surgical instruments.
- Maintained high segmentation performance (Dice 0.9728) comparable to the baseline (Dice 0.9723).
- Achieved parameter efficiency by updating only 4.48% of parameters with negligible latency increase.
Conclusions:
- Prototype-based open-set adaptation effectively rejects unseen surgical artifacts in AI-assisted colonoscopy.
- PolypSAM-Open preserves high polyp segmentation accuracy, crucial for clinical decision support.
- This approach offers a promising strategy for safer endoscopic workflows, pending clinical validation.
