Discriminative Joint-Feature Topic Model With Dual Constraints for WCE Classification.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence

Background:

• Wireless capsule endoscopy (WCE) generates vast image data for digestive tract examination.
• Robust image characterization is a significant challenge for automated diagnosis in WCE.
• Existing methods struggle with accurate classification of multiple abnormalities in WCE images.

Purpose of the Study:

• To develop a novel discriminative joint-feature topic model (DJTM) for classifying multiple abnormalities in WCE images.
• To improve the accuracy and robustness of computer-aided diagnosis in WCE.
• To characterize WCE images using distributions of latent semantic topics.

Main Methods:

• Proposed a joint-feature probabilistic latent semantic analysis (PLSA) model integrating color and texture descriptors.
• Introduced dual constraints: visual word importance and local image manifold, into the PLSA model.
• Characterized each WCE image by its distribution of learned latent semantic topics.

Main Results:

• The proposed DJTM achieved an excellent overall recognition accuracy of 90.78%.
• The model effectively integrates color and texture information for improved classification.
• DJTM demonstrated superior performance compared to existing multiple abnormalities classification methods for WCE images.

Conclusions:

• The DJTM offers a robust and accurate approach for classifying multiple abnormalities in WCE images.
• This method enhances computer-aided diagnosis by leveraging joint features and topic modeling.
• The DJTM represents a significant advancement in the automated analysis of WCE data.