Pediatric pneumonia is a critical respiratory infec tion that requires rapid and accurate diagnosis. While deep learning has shown promise in analyzing chest X-rays, dis tinguishing between bacterial and viral pneumonia remains a significant clinical challenge due to subtle visual differences. In this paper, the Multiscale Hybrid CNN-Transformer (MHC Transformer) is proposed, a novel architecture that synergizes local feature extraction using a pre-trained TorchXRayVision DenseNet-121 with global contextual modeling via a lightweight Vision Transformer (ViT) encoder. The extracted features are fused using a Multi-Head Cross-Attention mechanism. To address data leakage, a strict patient-wise split (70/15/15) is implemented on the Kermany dataset (5,856 unique images). The prepro cessing pipeline integrates Contrast Limited Adaptive Histogram Equalization (CLAHE) and TorchXRayVision normalization. To mitigate class imbalance and improve viral pneumonia recall, a custom Multi-class Focal Loss with label smoothing and Mixup augmentation is utilized. Evaluated using 8-pass Test Time Augmentation (TTA) and a soft-voting ensemble strategy, the proposed model achieved a superior overall accuracy of 94%, outperforming existing benchmarks. The proposed framework demonstrates robust clinical generalizability and incorporates Grad-CAM for interpretability.