Introducing a Deep Neural Network Model with Practical Implementation for Polyp Detection in Colonoscopy Videos.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Deep learning is increasingly utilized in computer-assisted minimally invasive surgery.
• Applications in colonoscopy include image analysis, operations analysis, skill evaluation, and automation.
• Deep learning for surgical image analysis is key for early gastrointestinal lesion detection and cancer treatment.

Purpose of the Study:

• To develop a simple and accurate deep learning model for polyp detection in colonoscopy.
• To address limited labeled data challenges using transfer learning and multi-task learning.
• To achieve both polyp classification and bounding box detection.

Main Methods:

• A deep neural network structure was proposed.
• Transfer learning and multi-task learning were employed.
• The model was trained and validated on KVASIR-SEG, CVC-CLINIC, and LDPolyp datasets, including custom non-polyp images.

Main Results:

• Achieved 100% accuracy in polyp/non-polyp classification.
• Reached 86% accuracy in bounding box detection.
• Demonstrated rapid processing time of 0.01 seconds for real-time application.

Conclusions:

• The deep learning model provides an efficient, accurate, and cost-effective solution for real-time colonoscopic polyp detection.
• The model's performance on benchmark datasets indicates potential for clinical deployment.
• Aids in early cancer diagnosis and treatment planning.