White-Light Endoscopic Colorectal Lesion Detection Based on Improved YOLOv5.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• White-light endoscopy faces challenges in detecting colorectal lesions, leading to missed and false detections.
• Improving the accuracy of colorectal cancer detection is crucial for patient outcomes.

Purpose of the Study:

• To develop a real-time lesion diagnosis model (YOLOv5x-CG) to enhance colorectal lesion detection rates.
• To subdivide colorectal lesions into three categories: micropolyps, adenomas, and cancer for precise diagnosis.

Main Methods:

• Utilized an improved YOLOv5 architecture (YOLOv5x-CG) for real-time lesion diagnosis.
• Implemented Mosaic data enhancement to improve the detection of small polyps.
• Incorporated coordinate attention (CA) mechanism for effective feature extraction.
• Employed the Ghost module to optimize feature generation and accelerate detection.

Main Results:

• The YOLOv5x-CG model demonstrated rapid and accurate lesion detection capabilities.
• Achieved high Average Precision (AP) values: 0.923 for polyps, 0.955 for adenomas, and 0.87 for cancer.
• Attained a mean Average Precision (mAP@50) of 0.916, indicating strong overall performance.

Conclusions:

• The proposed YOLOv5x-CG model significantly improves lesion detection rates in colorectal cancer patients.
• The integration of advanced techniques enhances the model's ability to differentiate and detect various colorectal pathologies.
• This AI-driven approach offers a promising tool for more effective endoscopic diagnosis.