Semi-supervised WCE image classification with adaptive aggregated attention.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate classification of Wireless Capsule Endoscopy (WCE) images is vital for early gastrointestinal (GI) cancer diagnosis.
• Challenges include limited annotated data, high intra-class variance, and inter-class similarities.
• Existing methods struggle with the complexities of WCE image analysis.

Purpose of the Study:

• To develop a novel semi-supervised learning method for automatic WCE image classification.
• To address the limitations of small datasets and complex image variations in WCE analysis.
• To enhance the accuracy and robustness of abnormality detection in the GI tract.

Main Methods:

• A semi-supervised learning framework incorporating an Adaptive Aggregated Attention (AAA) module.
• A preprocessing strategy using deformation fields to clean WCE images.
• A synergic network with two branches: abnormal region estimation and abnormal information distillation.
• Joint optimization using discriminative angular (DA) loss and Jensen-Shannon (JS) divergence loss on labeled and unlabeled data.

Main Results:

• The proposed method achieved 93.17% overall accuracy in a fourfold cross-validation on the CAD-CAP WCE dataset.
• Demonstrated effectiveness in overcoming challenges of limited data and image variability.
• The AAA module successfully captured global dependencies and contextual information.

Conclusions:

• The novel semi-supervised learning method with AAA module is effective for WCE image classification.
• The approach offers a promising solution for early GI cancer diagnosis using WCE.
• The developed technique improves the accuracy and reliability of automated WCE analysis.