Adaptive Detector-Verifier Framework for Zero-Shot Polyp Detection in Open-World Settings

2025-12-13

AI Summary

In open-world colonoscopy, zero-shot polyp detection suffers severe performance degradation due to illumination variations, motion blur, and occlusions. To address this, we propose a two-stage detector–verifier adaptive framework. Our method introduces a vision-language model (VLM)-guided frame-level dynamic confidence thresholding mechanism and a clinically motivated asymmetric cost-sensitive reward function—designed to penalize false negatives more heavily—to guide verifier refinement. We further employ Grouped Relative Policy Optimization (GRPO), a reinforcement learning algorithm, for verifier fine-tuning. Evaluated on YOLOv11 with a synthetic degradation benchmark, our approach achieves 14–22 percentage-point improvements in recall under zero-shot degradation settings on CVC-ClinicDB and Kvasir-SEG, while maintaining precision fluctuations within ±1.7%. This significantly mitigates false-negative risk in clinical deployment.

Technology Category

• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search
• Machine Learning: Calibration & Uncertainty Quantification

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Polyp detectors trained on clean datasets often underperform in real-world endoscopy, where illumination changes, motion blur, and occlusions degrade image quality. Existing approaches struggle with the domain gap between controlled laboratory conditions and clinical practice, where adverse imaging conditions are prevalent. In this work, we propose AdaptiveDetector, a novel two-stage detector-verifier framework comprising a YOLOv11 detector with a vision-language model (VLM) verifier. The detector adaptively adjusts per-frame confidence thresholds under VLM guidance, while the verifier is fine-tuned with Group Relative Policy Optimization (GRPO) using an asymmetric, cost-sensitive reward function specifically designed to discourage missed detections -- a critical clinical requirement. To enable realistic assessment under challenging conditions, we construct a comprehensive synthetic testbed by systematically degrading clean datasets with adverse conditions commonly encountered in clinical practice, providing a rigorous benchmark for zero-shot evaluation. Extensive zero-shot evaluation on synthetically degraded CVC-ClinicDB and Kvasir-SEG images demonstrates that our approach improves recall by 14 to 22 percentage points over YOLO alone, while precision remains within 0.7 points below to 1.7 points above the baseline. This combination of adaptive thresholding and cost-sensitive reinforcement learning achieves clinically aligned, open-world polyp detection with substantially fewer false negatives, thereby reducing the risk of missed precancerous polyps and improving patient outcomes.

Problem

• Polyp detectors fail in real-world endoscopy due to image degradation
• Domain gap exists between clean lab data and adverse clinical conditions
• Missed detections of precancerous polyps pose critical clinical risks

Innovation

• Adaptive thresholding guided by vision-language model for detection
• Cost-sensitive reinforcement learning fine-tunes verifier with GRPO
• Synthetic testbed simulates adverse clinical conditions for evaluation