CRH-YOLO for precise and efficient detection of gastrointestinal polyps

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Gastrointestinal polyps are critical indicators of digestive system diseases, necessitating early detection.
• Current polyp detection methods face challenges with uncertainty and scale, impacting patient health.
• Timely endoscopic interventions reduce malignancy risk but require improved detection tools.

Purpose of the Study:

• To propose a novel single-stage method for enhanced gastrointestinal polyp detection.
• To improve the accuracy and efficiency of polyp identification in endoscopic images.
• To address limitations in current methods regarding scale and uncertainty in polyp detection.

Main Methods:

• Development of a novel single-stage polyp detection model named CRH-YOLO.
• Integration of the Contextual Relationship Feature Enhancement Module (CRFEM) for improved target perception.
• Implementation of the Residual Squeeze-and-Excitation Pyramid Pooling Fusion (RSPPF) for multi-scale feature fusion.
• Modification of the original model by removing a detection head and adding a high-dimensional feature compensation structure.

Main Results:

• CRH-YOLO achieved 88.8% Precision, 86.0% Recall, and 90.7% map@.5 on the Kvasir-seg dataset.
• The model outperformed mainstream detection models like YOLOv8n, with a 2.4% improvement in map@.5.
• CRH-YOLO demonstrated superior performance in detecting small or subtle polyps.

Conclusions:

• CRH-YOLO offers an effective solution for the early detection and prediction of gastrointestinal polyps.
• The proposed model enhances contextual information perception and multi-scale feature fusion for improved detection.
• The optimized model structure reduces parameters while maintaining high detection performance, especially for challenging cases.