Hybrid Feature-Fusion Model Combining GhostNet and MobileNetV2 for Automated Pneumonia Detection

Abstract

Pneumonia remains a significant global health concern, especially in regions with limited medicalresources, underscoring the need for accurate, efficient, and interpretable diagnostic solutions. Themodel leverages GhostNet’s efficient feature extraction and MobileNetV2’s lightweight precision.Fusion is performed after the final convolutional blocks of GhostNet and MobileNetV2, where feature maps are aligned using adaptive pooling and merged through channel-wise con-catenation. Training was conducted on a publicly available pediatric chest X-ray dataset comprising 5,872 images from the Guangzhou Women and Children’s Medical Center. A patient-level split of 70% for training, 15% for validation, and 15% for testing was used, ensuring no data leakage across subsets. Although cross-validation was not applied, generalizability was assessed on an external adult dataset from Indiana University (Open-i), with the model achieving 85% test accuracy and 87% validation accuracy. External validation was conducted on the Indiana University Open-i dataset using the same preprocessing and inference pipeline as the internal dataset to ensure consistent cross-domain evaluation. Benchmarking against state-of-the-art models including DenseNet121, Effi-cientNetV2L, ResNet50, and VGG16 demonstrated that the proposed hybrid model achieves competitive or superior accuracy while maintaining substantially lower computational cost. On the internal test set, the proposed method attained 9.47% accuracy, 99.60% precision, 95.64% recall, and a 97.56% F1-score. Training and validation loss curves showed minimal divergence, and Grad-CAM visualizations offered interpretability by highlighting salient lung regions influencing predictions. As only a single train–validation–test split was used, confidence intervals, statistical significance tests, and variance across multiple runs were not calculated, representing a limitation in the robustness of the reported results. The lightweight and adaptable nature of the model makes it particularly suitable for real-world deployment in resource-constrained healthcare environments. Future work will focus on expanding the dataset, adopting k-fold cross-validation, integrating continual learning strategies, conducting subgroup and fairness analyses, and exploring explainable AI tools to further enhance clinical applicability and trust.