Integrating Local and Global Representation Learning for Pediatric Pneumonia Detection: A Hybrid CNN-Transformer...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Pneumonia is a major global cause of childhood illness and death.
• Interpreting pediatric chest X-rays is difficult due to anatomical variations and subtle findings.
• Accurate AI detection of pediatric pneumonia is crucial for timely treatment.

Purpose of the Study:

• To evaluate various Convolutional Neural Network (CNN)-Transformer ensemble strategies for pediatric pneumonia detection.
• To combine local and global image features for improved diagnostic accuracy.
• To assess the performance and interpretability of different ensemble methods.

Main Methods:

• Utilized the Pediatric Pneumonia Chest X-ray dataset (5856 radiographs).
• Employed EfficientNetV2-S for local features and Swin Transformer-T for global relationships.
• Compared soft voting, weighted voting, and stacking ensemble techniques.
• Applied image preprocessing, data augmentation, and Grad-CAM for interpretation.

Main Results:

• Ensemble learning outperformed individual models in pneumonia detection.
• Soft Ensemble achieved the highest accuracy (96.96%) and F1-score (97.59%) on the hold-out set.
• Hybrid CNN-Transformer Stacking yielded the highest sensitivity (99.49%) and fewest false negatives.
• Grad-CAM confirmed complementary feature extraction by CNN and Transformer models.

Conclusions:

• Different ensemble strategies offer unique benefits for pediatric pneumonia detection.
• Soft voting provides excellent overall performance.
• Stacking excels in minimizing false negatives and maximizing sensitivity, showing promise for AI-assisted screening.