Deep transfer learning approaches for bleeding detection in endoscopy images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Wireless capsule endoscopy generates vast image data, making manual analysis by physicians time-consuming.
• Computer-aided diagnosis (CAD) systems are crucial for automating the analysis of capsule endoscopy images.
• Deep learning methods have significantly advanced the classification of endoscopic images.

Purpose of the Study:

• To develop an expert system for automatic classification of wireless capsule endoscopy images.
• To improve the efficiency and accuracy of detecting bleeding and non-bleeding regions.
• To evaluate the performance of a novel deep learning-based feature extraction and fusion approach.

Main Methods:

• Employed three pre-trained deep convolutional neural networks (VGG19, InceptionV3, ResNet50) for feature extraction.
• Utilized the Minimum Redundancy Maximum Relevance (mRMR) method for feature selection and fusion.
• Applied supervised machine learning algorithms, including Support Vector Machine (SVM), for image classification.

Main Results:

• The proposed architecture achieved high accuracy in detecting bleeding regions, averaging 97.65% and 95.70% on benchmark datasets.
• The system outperformed single deep learning architectures in classification tasks.
• Mean value pooling as a fusion rule combined with SVM classifier yielded optimal accuracy and training time.

Conclusions:

• The developed expert system effectively classifies wireless capsule endoscopy images, significantly improving bleeding detection accuracy.
• Feature fusion techniques enhance the performance of deep learning models for medical image analysis.
• This approach offers a promising solution for efficient and accurate analysis of capsule endoscopy data.