Gastrointestinal tract disease classification from wireless capsule endoscopy images based on deep learning...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Cancer is a leading cause of death globally.
• Diagnosing gastrointestinal tract (GIT) abnormalities via endoscopy faces challenges due to interobserver variability and expertise.
• Subtle lesions can be missed, delaying early cancer diagnosis.

Purpose of the Study:

• To develop an advanced computer-aided diagnosis (CAD) system for early detection and classification of GIT diseases.
• To enhance diagnostic accuracy and overcome limitations of manual endoscopic interpretation.
• To leverage deep learning and feature optimization for improved GIT abnormality identification.

Main Methods:

• A novel CAD system combining feature fusion of modified deep learning models (ResNet18, ResNet50 with EFP layers) and optimal feature selection.
• Utilized Kvasir V1, Kvasir V2, and Hyperkvasir datasets for training and validation.
• Employed contrast enhancement via top-bottom filtering, entropic field propagation (EFP) layers, and Newton-Raphson Marine Predator Optimization (NRMPO) for feature refinement.

Main Results:

• The proposed CAD system achieved high accuracies: 99.0% on Kvasir V1, 89.6% on Kvasir V2, and 82.7% on Hyperkvasir.
• Feature fusion using a novel mean threshold-based approach improved classification performance.
• Ablation studies confirmed the efficacy of the intermediate steps in achieving the reported accuracies.

Conclusions:

• The developed CAD system demonstrates superior performance compared to state-of-the-art techniques.
• The method offers improved accuracy and precision rates for GIT disease diagnosis.
• This approach holds significant potential for early and reliable detection of gastrointestinal abnormalities.