<![CDATA[Pediatric pneumonia diagnosis through chest X-ray analysis is complicated by subtle radiographic patterns and diagnostic subjectivity. A deep learning architecture integrating transfer learning with EfficientNetB4 as a feature extraction backbone is proposed, enhanced by a supplementary 3×3 convolutional layer (ReLU activation) and global average pooling to preserve localized pathological features. The dataset comprises 5,863 pediatric anterior-posterior chest radiographs curated from Guangzhou Women and Children’s Medical Center, rigorously validated by three board-certified radiologists to ensure diagnostic fidelity. Stratified sampling allocated 80% for training, 10% for validation, and 10% for testing, with stochastic augmentation (rotation: ±5°, width/height shift: ±10%, shear: 20%, horizontal flip) addressing class imbalance and enhancing model generalizability. Training employed Adam optimization (initial learning rate: 0.001) with binary cross-entropy loss, dynamically modulated via ReduceLROnPlateau (factor: 0.3, patience: 3). Independent test evaluation yielded 97.7% accuracy (95% CI: 96.8–98.5%), AUC-ROC of 0.9954, and F1-scores of 0.9842 (pneumonia) and 0.9573 (normal), supported by a Matthews correlation coefficient (MCC) of 0.9416 and Cohen’s Kappa of 0.9416. Precision-recall analysis demonstrated a 98.4% positive predictive value for pneumonia identification. The architecture’s robustness to imaging variability and high diagnostic precision positions it as a scalable triage tool in low-resource healthcare settings, potentially reducing diagnostic latency and improving pediatric outcomes.]]>
