Effective high-to-low-level feature aggregation network for endoscopic image classification.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Endoscopic image classification is crucial for diagnosing intestinal diseases.
• Current Convolutional Neural Network (CNN) methods often overlook vital low-level features, focusing only on abstract high-level features.
• Low-level features are significant for accurate diagnosis in gastroenterology.

Purpose of the Study:

• To develop an improved endoscopic image classification method by effectively utilizing both high-level and low-level features.
• To enhance diagnostic accuracy for endoscopists through better image analysis.
• To address the limitations of existing CNN-based approaches in endoscopic image classification.

Main Methods:

• A novel two-stream network architecture was proposed for endoscopic image classification.
• The backbone stream extracts high-level features, while a fusion stream extracts low-level features using a bottom-up multi-scale gradual integration (BMGI) method.
• Top-down attention learning modules refine BMGI input, and a novel correction loss function clarifies feature relationships.

Main Results:

• The proposed framework achieved an overall classification accuracy of 97.33% on the KVASIR dataset.
• A Kappa coefficient of 95.25% was obtained, indicating high agreement.
• The evaluation metrics showed at least a 2% improvement compared to existing state-of-the-art models.

Conclusions:

• The developed two-stream network effectively fuses high-level and low-level features for superior endoscopic image classification.
• The proposed method outperforms current classification approaches by better representing endoscopic images.
• The novel correction loss regularizes feature consistency, reducing intra-class distances and improving label prediction accuracy.