Efficient and Accurate Pneumonia Detection Using a Novel Multi-Scale Transformer Approach.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer-Aided Diagnosis

Background:

• Pneumonia is a major global health concern, causing significant morbidity and mortality.
• Chest X-rays are crucial for pneumonia diagnosis but face interpretation challenges due to image variability.
• Automated systems can improve diagnostic consistency and support clinical decisions in pneumonia detection.

Purpose of the Study:

• To develop a novel multi-scale transformer approach for integrated lung segmentation and pneumonia classification.
• To enhance the accuracy and efficiency of automated pneumonia detection from chest X-rays.
• To create a computationally efficient model suitable for resource-limited clinical settings.

Main Methods:

• A lightweight transformer-enhanced TransUNet was utilized for precise lung segmentation (95.68% Dice score).
• Pre-trained ResNet models (ResNet-50, ResNet-101) extracted multi-scale features for classification.
• A convolutional Residual Attention Module and a modified transformer processed features to improve pneumonia detection.

Main Results:

• The proposed method achieved 93.75% accuracy on the Kermany dataset and 96.04% accuracy on the Cohen dataset.
• The lung segmentation component demonstrated high performance with fewer parameters than traditional transformers.
• The integrated approach outperformed existing methods in pneumonia detection accuracy and computational efficiency.

Conclusions:

• The novel multi-scale transformer approach offers a robust and efficient solution for pneumonia detection.
• The unified framework for segmentation and classification enhances diagnostic reliability in medical imaging.
• This AI-driven method shows promise for improving patient outcomes and supporting clinical workflows in pneumonia diagnosis.