Deep Learning Based Disease Classification Using Chest X-Ray Images

Accurate and efficient detection of chest infections diseases with X-ray images employing deep learning is crucial for the early detection of pulmonary conditions such as COVID-19, viral pneumonia, and lung opacity. In this study, a comparative evaluation of two fine-tuning strategies of the ResNet-50 deep convolutional neural network was conducted to investigate the trade-offs between classification performance and computational efficiency. The first approach involved full fine-tuning of all network layers, while the second approach retrained only the classifier layers, keeping the feature extractor frozen. Both models were trained and evaluated using the COVID-19 Radiography Database, evaluation through accuracy, F1-score, confusion matrices, and training time analysis. The results demonstrated that the fully fine-tuned model achieved higher classification accuracy (96.4%) and F1-score (0.9621), particularly improving the detection of critical classes such as COVID-19 and lung opacity. However, this came at the cost of increased computational resources, with 23.5 million trainable parameters and longer training times compared to the partial fine-tuning model, which offered a lightweight alternative with 2.05 million parameters and 44% faster training. This study highlights the accuracy-efficiency trade-off inherent in deep learning fine-tuning strategies and provides practical recommendations for selecting appropriate models based on deployment scenarios. Future work will explore scalable and resource-efficient deep learning approaches, including model pruning, knowledge distillation, and real-world clinical validations, to enhance the applicability of such systems in healthcare settings.