Utilizing Deep Convolutional Neural Networks and Hybrid Classification for Gastrointestinal Disease Diagnosis from...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Gastroenterology
• Deep Learning for Diagnostic Systems

Background:

• Wireless Capsule Endoscopy (WCE) provides non-invasive gastrointestinal (GI) visualization but generates extensive data, making manual review time-consuming and prone to errors.
• Accurate detection of GI lesions requires high sensitivity and Negative Predictive Value (NPV) from diagnostic systems.
• Individual deep learning models struggle with the complexity of classifying diverse GI lesions in WCE frames.

Purpose of the Study:

• To develop a reliable expert diagnostic system for WCE frame analysis using a hybrid classification approach.
• To enhance the accuracy of classifying prevalent GI lesions by overcoming limitations of single deep learning models.
• To improve diagnostic performance metrics like sensitivity and NPV for gastroenterologists.

Main Methods:

• Trained Deep Convolutional Neural Networks (DCNNs) on balanced and unbalanced WCE datasets.
• Implemented ensemble learning techniques to create hybrid classifiers from DCNNs using model scoring.
• Utilized class scoring to refine decision boundaries within hybrid classifiers for improved multi-class classification.

Main Results:

• The VG_BFCG model, based on VGG16, achieved a recall of 0.952 and NPV of 0.977.
• Hybrid classifiers using class scoring demonstrated superior performance, reaching a recall of 0.988 and NPV of 1.00.
• An unbalanced dataset with more Angiectasia frames positively impacted classification metrics across all models.

Conclusions:

• Class scoring is crucial for enhancing multi-class hybrid classification performance in WCE analysis.
• The developed hybrid system significantly improves diagnostic accuracy, offering high sensitivity and NPV.
• This AI-driven approach aids gastroenterologists in more efficient and accurate WCE interpretation.