WCE-AGDN

WCE-AGDN enhances detection and classification of small lesions in wireless capsule endoscopy (WCE) images of the gastrointestinal (GI) tract by combining attention-guided deformation, Third-order Long-range Feature Aggregation (TLFA), and a Deformation based Attention Consistency (DAC) loss to improve diagnostic accuracy.

Key Features:

• Two-branch convolutional neural network: A dual-branch CNN architecture separates attention-map generation and amplified-input processing to improve lesion representation.
• Attention-guided region amplification: Attention maps from the first branch guide amplification of regions of interest in the second branch's input to enhance small lesion visibility.
• Third-order Long-range Feature Aggregation (TLFA): TLFA modules capture long-range dependencies and aggregate contextual features across the image for stronger feature representation.
• Deformation based Attention Consistency (DAC) loss: DAC enforces consistency between the two branches' attention maps to refine attention and promote mutual enhancement.
• Global feature fusion for classification: Global feature embeddings from both branches are fused to predict image labels.
• Performance on public WCE datasets: Reported overall classification accuracy of 91.29% on two public WCE datasets.
• Robustness to small lesions and background interference: Specifically addresses detection challenges posed by small lesions and background noise in WCE images.

Scientific Applications:

• Lesion detection in WCE: Enhances representation and detection of small gastrointestinal lesions in wireless capsule endoscopy images.
• Image-level classification of WCE frames: Classifies WCE images to support diagnostic labeling of GI tract findings.
• Noninvasive gastrointestinal diagnostic imaging: Improves diagnostic accuracy in noninvasive visualization of the entire GI tract using WCE.

Methodology:

Two-branch convolutional neural network where the first branch generates attention maps that guide amplification of input regions for the second branch; Third-order Long-range Feature Aggregation (TLFA) modules capture long-range dependencies and aggregate contextual features; a Deformation based Attention Consistency (DAC) loss enforces consistency between branch attention maps; global feature embeddings from both branches are fused to predict image labels.