Pneumonia remains a major global health concern, particularly in low-resource regions where access to expert radiologists is limited. Deep learning has shown strong potential for chest X-ray (CXR) analysis, yet many high-performing models rely on large backbones and computationally expensive attention modules, limiting deployment on portable or embedded devices. Mobile-oriented architectures such as MobileNetV3 and lightweight attention mechanisms like SE and CBAM have improved efficiency, but achieving high diagnostic accuracy with extremely small model complexity remains challenging. We propose CSANet++, a substantially improved and lightweight extension of our previously accepted CSANet framework. CSANet++ introduces four key innovations: a MobileNetV3Small backbone for efficient feature extraction, a Lite Channel–Spectral Attention (LiteCSA) module using depthwise spectral filtering and channel squeeze, a Lite Spatial Attention (LiteSPA) module leveraging anatomical priors of upper–middle–lower lung zones, and a Gated Fusion mechanism that dynamically balances spectral and spatial cues. Additionally, convolution–batch normalization fusion is employed to reduce inference latency. Despite its compact size of 1.15M parameters and 58.5M MACs, CSANet++ achieves 97.27% accuracy, 97.79% precision, 98.48% recall, 98.14% F1-score, and 0.9968 AUC on the pediatric CXR dataset. These results outperform heavier CNN baselines while maintaining real-time performance (0.0049 s/img), demonstrating the model’s suitability for on-device pneumonia screening.