An improved deep learning approach and its applications on colonic polyp images detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colonic polyps can develop into colon cancer if not treated early.
• Colonoscopy accuracy is limited by operator experience and fatigue.
• Early detection of colonic polyps is crucial for preventing colon cancer.

Purpose of the Study:

• To improve the accuracy and efficiency of colonic polyp detection during colonoscopy.
• To develop a deep learning-based computer-aided diagnosis system for colonoscopy.
• To assist endoscopists in identifying potentially overlooked polyps in real-time.

Main Methods:

• Utilized colonoscopy images from Hunan children's hospital for dataset creation.
• Applied deep learning image classification, specifically VGGNets and ResNets.
• Proposed improved VGGNets-GAP and ResNets-GAP models incorporating global average pooling.

Main Results:

• Achieved classification accuracies exceeding 98% across all models.
• Demonstrated high sensitivity (TPR > 96%) and specificity (TNR > 98%).
• VGGNets-GAP models showed high accuracy with significantly reduced parameters compared to original VGGNets.

Conclusions:

• The deep learning approach effectively aids in the automatic detection of colonic polyps.
• The proposed method enhances detection accuracy and reduces missed diagnoses.
• The VGGNets-GAP model is lightweight, decreasing memory consumption and improving efficiency.