A hybrid explainable ensemble transformer encoder for pneumonia identification from chest X-ray images

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Pneumonia detection relies on chest X-rays, often requiring advanced feature extraction due to image quality limitations.
• Accurate and early pneumonia diagnosis is crucial for effective patient treatment and management.
• Existing AI models for pneumonia detection face challenges in interpretability and performance.

Purpose of the Study:

• To propose a hybrid explainable deep learning framework for accurate pneumonia identification using chest X-ray images.
• To enhance the interpretability of deep learning models in medical diagnostics.
• To achieve superior performance in pneumonia classification compared to existing methods.

Main Methods:

• A hybrid framework fusing ensemble convolutional networks (Ensembles A and B) with a Transformer Encoder.
• Ensemble learning extracts robust features from X-ray images.
• Transformer Encoder with self-attention and MLP for classification; saliency maps provide visual explanations.

Main Results:

• The hybrid model achieved 99.21% accuracy and F1-score for binary classification and 98.19% accuracy/97.29% F1-score for multi-class classification.
• Ensemble A and B demonstrated strong performance in both binary and multi-class scenarios, with accuracies ranging from 96.44% to 97.22%.
• The framework provides explainable saliency maps highlighting critical regions in X-ray images.

Conclusions:

• The proposed hybrid deep learning framework offers promising and explainable pneumonia identification performance.
• It surpasses individual, ensemble, and state-of-the-art AI models in pneumonia detection accuracy and interpretability.
• The framework's code is publicly available for further research and application.