A hybrid XAI-driven deep learning framework for robust GI tract disease diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Gastrointestinal (GI) diseases pose significant diagnostic challenges, with misclassification leading to adverse patient outcomes.
• Current AI methods for medical image analysis often exhibit high false-negative rates, compromising critical diagnoses.
• Accurate and timely diagnosis of stomach diseases is crucial for effective treatment and patient quality of life.

Purpose of the Study:

• To develop a hybrid deep learning (DL) and explainable artificial intelligence (XAI) model for enhanced accuracy in diagnosing GI tract diseases from endoscopic images.
• To improve diagnostic accuracy and model interpretability by integrating Swin Transformer with Deep Convolutional Neural Networks (DCNNs).
• To minimize false negatives in GI disease evaluation through stacked machine learning classifiers, meta-loss, and Grad-CAM visualization.

Main Methods:

• A hybrid DL model combining Swin Transformer with DCNNs (EfficientNet-B3, ResNet-50) was utilized for feature extraction.
• Stacked machine learning classifiers were employed with meta-loss and Grad-CAM (an XAI technique) to reduce false negatives.
• The model was trained and evaluated on endoscopic images for the classification of gastrointestinal disorders.

Main Results:

• The proposed model achieved an overall accuracy of 93.79% in classifying gastrointestinal tract diseases.
• Significant improvements were observed in class-wise performance metrics, including precision, recall, and F1-score.
• The false-negative rate was substantially reduced, indicating enhanced diagnostic reliability.

Conclusions:

• The synergistic DL and XAI approach offers a promising method for improving the accuracy and interpretability of AI-driven GI disease diagnosis.
• Grad-CAM provides visual explanations, making AI predictions more accessible and understandable for medical professionals.
• This study paves the way for earlier diagnosis of GI diseases with reduced human error, supporting clinical decision-making.