EVAD-YOLO: An endoscopic video anomaly detection based on improved YOLOV11.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Gastrointestinal diseases are increasing, necessitating early detection.
• Automated analysis of endoscopy videos is crucial for clinical screening.
• Accurate identification of complex gastrointestinal lesions remains a challenge.

Purpose of the Study:

• To propose an improved YOLOv11-based model for detecting gastrointestinal lesions in endoscopic videos.
• To enhance the accuracy and robustness of lesion detection, including gastric ulcers and cancer.
• To provide a reliable tool for clinical-assisted endoscopic diagnosis.

Main Methods:

• Developed an Endoscopic Video Anomaly Detection based on improved YOLOV11 (EVAD-YOLO).
• Introduced a Residual Global Expansion Attention (RGEA) module for enhanced global contextual perception.
• Designed an Enhanced Multi-scale Fusion (EMSF) module for integrating features across scales.
• Utilized a mixed endoscopic dataset including polyps, gastric ulcers, and early gastric cancers.

Main Results:

• EVAD-YOLO demonstrated superior performance in detecting gastrointestinal lesions.
• Achieved 90.4% precision, 84.3% recall, and 90.4% mAP50.
• The model showed strong robustness for lesions with complex shapes, color variations, and varying sizes.

Conclusions:

• EVAD-YOLO offers a robust and accurate solution for automated gastrointestinal lesion detection.
• The proposed RGEA and EMSF modules significantly improve detection capabilities.
• The method shows strong potential for reliable clinical-assisted endoscopic diagnosis.