Computer-Aided Colon Polyp Detection on High Resolution Colonoscopy Using Transfer Learning Techniques.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colonoscopies are crucial for colorectal cancer prevention by detecting and removing polyps.
• Conventional colonoscopy has a significant polyp miss rate (up to 26%), highlighting the need for improved detection methods.
• Existing automated polyp detection systems often lack real-time capabilities due to computational limitations.

Purpose of the Study:

• To develop and evaluate an automated system for real-time colon polyp detection and classification using deep learning.
• To address the limitations of current systems in terms of speed and accuracy for clinical application.

Main Methods:

• Utilized convolutional neural network (CNN) transfer learning for image pattern recognition in real-time colonoscopy videos.
• Trained multiple multi-class classifiers and a Faster R-CNN detector based on the Inception v2 model.
• Evaluated model performance using mean Average Precision (mAP).

Main Results:

• The Faster R-CNN detector achieved a mAP of 77%, a substantial improvement over multi-class classifiers (38%-49%).
• The developed model demonstrated high accuracy in detecting colon polyps in real-time.
• Polyp type classification accuracy requires further enhancement.

Conclusions:

• CNN transfer learning offers a promising approach for enhancing real-time colon polyp detection accuracy.
• The Faster R-CNN model shows significant potential for improving colonoscopy outcomes.
• Further research is needed to refine polyp classification capabilities for comprehensive diagnostic support.