Optimization of an efficient net-based transfer learning model for automated pneumonia detection from chest x-ray images

Abstract

Pneumonia continues to be a major contributor to respiratory-related mortality worldwide, particularly within sub-Saharan regions where access to specialized radiological expertise remains limited. This study focuses on the design and optimization of an EfficientNet-based transfer learning model for the automated detection of pneumonia from chest X-ray images. The model was trained and validated using the publicly available Chest X-Ray (Pneumonia) dataset from Kaggle, comprising 5,856 radiographic images. Experimental evaluations were conducted across four data partitioning schemes: 80:10:10, 70:15:15, 60:20:20, and 75:15:10. Pre-processing steps involved image normalization and augmentation, while fine-tuning was performed using the Binary Cross-Entropy loss function and the Adam optimizer over 25 training epochs. Model performance was assessed using standard evaluation metrics, including accuracy, precision, recall, F1-score, and the ROC-AUC. Results indicated stable performance across all splits, with the 60:20:20 partition achieving the highest performance, (i.e., 90.50% test accuracy, 94.97% F1-score, and 0.725 ROC-AUC). Furthermore, the model achieved high recall values (≥98%) across all experiments, demonstrating exceptional sensitivity in identifying pneumonia-positive cases. These outcomes underscore the robustness and scalability of the EfficientNet architecture for medical image classification, even under varying data distributions. The proposed model offers a computationally efficient and clinically relevant solution for early pneumonia detection, particularly in resource-limited healthcare environments. Future research will focus on enhancing model interpretability and performing cross-dataset validation to further advance its clinical integration and reliability.