Enhanced gastrointestinal disease classification using a convvit hybrid model on endoscopic images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Endoscopy enables visualization of the gastrointestinal tract for disease diagnosis.
• Accurate processing of endoscopic images is crucial for early detection and treatment of gastrointestinal conditions.
• Existing deep learning models have limitations in capturing both local and global features in endoscopic images.

Purpose of the Study:

• To develop a hybrid Convolutional Neural Network (CNN) and Vision Transformer (ViT) model, termed ConvViT.
• To enhance the classification accuracy and reliability of detecting pathologies in gastrointestinal endoscopic images.
• To leverage the complementary strengths of CNNs for local feature extraction and ViTs for global dependency modeling.

Main Methods:

• Developed a hybrid ConvViT model integrating CNN and ViT architectures.
• Compared the performance of ConvViT against established models: VGG-16, ResNet-50, Inception-V3, and ViT.
• Utilized a diverse dataset of gastrointestinal endoscopic images featuring ulcers, polyps, inflammation, bleeding, and normal anatomy.

Main Results:

• The ConvViT model demonstrated superior prediction performance compared to all benchmarked models.
• Achieved a classification accuracy of 95.87% in identifying various gastrointestinal pathologies.
• The hybrid approach effectively combined local spatial feature detection and long-range dependency modeling.

Conclusions:

• The hybrid ConvViT model offers improved accuracy and reliability for classifying gastrointestinal endoscopic images.
• This AI-driven approach holds significant potential for advancing early diagnosis of gastrointestinal diseases.
• ConvViT represents a promising advancement in medical image analysis for gastroenterology.