SIG-CFFNet: Structural Information-Guided Cascaded Feature Fusion Network for Gastrointestinal Anatomy...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate identification of gastrointestinal endoscopic anatomical structures is crucial for diagnosis.
• Endoscopic image quality issues and morphological similarities hinder accurate recognition.
• Existing methods struggle with low-quality and complex endoscopic images.

Purpose of the Study:

• To develop an advanced deep learning model for precise identification of gastrointestinal endoscopic anatomical structures.
• To enhance diagnostic accuracy and reduce missed detection rates in endoscopy.
• To address challenges posed by low-quality images and complex anatomical variations.

Main Methods:

• Proposed a Structural Information-Guided Cascaded Feature Fusion Network (SIG-CFFNet).
• Integrated Convolutional Neural Network (CNN) and Transformer features using anatomical prior knowledge.
• Employed Depthwise Over-parameterized Convolutional Layer (DO-Conv) for enhanced feature representation and efficiency.

Main Results:

• Achieved 73.47% accuracy for normal and 87.05% for pathological structures.
• Attained 99.61% accuracy on Kvasir-Capsule and 87.83% on HyperKvasir datasets.
• Demonstrated robust cross-domain performance (84.46% COVID19-CT, 80.21% ISIC2018) with competitive recall rates.

Conclusions:

• SIG-CFFNet significantly improves the identification of gastrointestinal endoscopic anatomical structures.
• The model shows high accuracy and robustness across various datasets and challenging conditions.
• The proposed network has strong clinical applicability and potential for real-world endoscopic diagnosis.