Hybrid deep learning framework based on EfficientViT for classification of gastrointestinal diseases

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Gastrointestinal (GI) diseases are a major global health concern, necessitating early and accurate diagnosis.
• Current diagnostic methods like endoscopy are time-consuming and rely heavily on physician interpretation.
• Advancements in artificial intelligence offer potential for improving diagnostic accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate a novel deep learning model, Efficient Vision Transformer (EfficientViT), for classifying eight types of GI diseases.
• To leverage the strengths of EfficientNetB0 and Vision Transformer (ViT) for enhanced feature extraction in GI endoscopic images.
• To provide a more reliable and accurate tool for clinicians in diagnosing GI diseases.

Main Methods:

• Proposed Efficient Vision Transformer (EfficientViT) model combining EfficientNetB0 and Vision Transformer (ViT).
• EfficientViT utilizes EfficientNetB0 for local texture and multi-scale feature capture, and ViT for global context recognition.
• A dual-block design optimizes the model by processing input for local details (EfficientNet) and global dependencies (encoder block) simultaneously.

Main Results:

• EfficientViT achieved an outstanding accuracy of 99.82% in classifying GI diseases.
• The model demonstrated superior performance compared to a MobileNetV2-based model (99.60% accuracy).
• EfficientViT exhibited excellent precision, recall, and F1 scores, outperforming existing methods.

Conclusions:

• EfficientViT offers a highly accurate and reliable AI-powered tool for diagnosing GI diseases from endoscopic images.
• The model's ability to capture both local and global features enhances diagnostic capabilities.
• This deep learning approach presents a promising advancement for clinical practice in gastroenterology.