Strategic Residual Connections: Enhancing Vision Transformers for Pediatric Pneumonia Classification

Pneumonia detection from chest radiographs remains a critical challenge in pediatric healthcare, where accurate and rapid diagnosis can significantly impact patient outcomes. This study performs an in-depth assessment of Vision Transformer (ViT) systems designed to automatically detect pneumonia signatures in pediatric chest radiographic images. Two primary approaches are investigated: a Fine-Tuned Fine-Tuned Vanilla ViT model and a novel Custom ViT with Residual Convolutional units Modification. The experimental framework utilized a diverse dataset comprising 5,865 pediatric chest radiographs, ensuring robust validation of the proposed methodologies. The fine-tuned Fine-Tuned Vanilla ViT demonstrated strong baseline performance, achieving 95.35% accuracy in pneumonia classification. However, the Custom ViT architecture, enhanced with strategically placed residual convolutional units, exhibited superior performance across all metrics: 98.08% accuracy, 98.35% precision, and 99.10% recall, yielding an F1-score of 0.9872. The findings suggest that carefully designed ViT architectures, when combined with domain-specific enhancements, can serve as powerful tools for automated medical image analysis, enhancing both the efficacy and robustness of pneumonia detection in clinical practice.