Transfer Learning-Based COVID-19 Detection Using Chest X-Ray Images with a Pretrained Neural Network and Custom Classifier

The unfolding global trajectory of COVID-19 has made it abundantly clear that scalable, accurate diagnostic solutions, especially in resource-limited healthcare settings, are urgently required. The present study therefore proposes a novel, transfer learning-based framework for COVID-19 detection employing chest X-ray imagery, repurposing the Google Derm Foundation neural network, a model originally trained for dermatological classification. By integrating domain-specific feature extraction with a custom neural network classifier optimized for binary COVID-versus-non-COVID inference, it constitutes a hybrid architecture that transcends conventional approaches. The pipeline incorporates rigorous preprocessing techniques, stratified cross-validation, and exhaustive ablation studies to guarantee robustness and reliability. Key contributions include the validation of the Google Derm Foundation model's adaptability to pulmonary diagnostics and the systematic performance comparison of multiple network architectures. Empirical results reveal notable superiority: the Google derm-foundation achieved an accuracy of 94.39%, precision of 95.72%, and F1-score of 94.31%, outperforming traditional classifiers and pretrained convolutional networks such as VGG19 and ResNet50. Full visualization of the models, including their use of confusion matrices and accuracy curves, allows for additional insights into the model's behavior to be gleaned. The close inspection of erroneous negatives highlights the clinical consequences of making a mistaken diagnosis. This research advances AI-driven medical imaging by presenting a scalable, interpretable, and highperforming solution for COVID-19 detection. The next steps consist of data augmentation, class-imbalance mitigation, and domain-adaptation approaches aimed at making the technologies clinically even more useful.