An automated detection system for colonoscopy images using a dual encoder-decoder model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Detection

Background:

• Conventional computer-aided detection (CAD) systems for colonoscopy have limitations in sensitivity and specificity.
• Existing methods often rely on shape, texture, or temporal features, which are insufficient for accurate polyp detection.

Purpose of the Study:

• To develop a lightweight dual encoder-decoder convolutional neural network (CNN) model for automatic polyp feature extraction and detection in colonoscopy images.
• To achieve performance comparable to deeper models with a more efficient, shallow architecture.

Main Methods:

• A dual CNN architecture comprising a hetero-encoder and an auto-encoder was proposed.
• The hetero-encoder generates corrupted labeled images to reduce reliance on large training datasets.
• The auto-encoder learns features from noisy images to enhance classification discriminative power.

Main Results:

• The proposed model demonstrated improved performance compared to a state-of-the-art detection model.
• Key performance metrics included Jaccard index, DICE similarity score, and geometric measures.
• Improvements were attributed to reduced false positives by the auto-encoder and effective noisy image generation by the hetero-encoder.

Conclusions:

• The proposed lightweight dual CNN model offers an effective approach for polyp detection in colonoscopy.
• The method shows potential for improving the accuracy and efficiency of computer-aided detection systems.
• This architecture addresses the need for robust polyp detection with reduced training data requirements.