Exploring Deep Convolutional Neural Networks: A Grad-CAM Enhanced Comparative Study for Automated COVID-19 Diagnosis from Chest X-Ray Images

The emergence of COVID-19 in late 2019 sparked a global health crisis. Identifying infections quickly is key to providing proper care and limiting transmission. The Reverse Transcription-Polymerase Chain Reaction (RT-PCR) test, a widely recognized primary diagnostic method for detecting COVID-19, faces challenges associated with cost and compatibility in resource-constrained environments. This issue is resolved by chest X-ray imaging since it is non-invasive and inexpensive. In this domain, research often utilizes pre-trained models on relatively small datasets of chest X-ray images, limiting the potential for optimal generalization and accuracy achievable with larger training sets. Also, these research works limit the exploration of crucial information on salient regions within CXR images, constraining the depth of analysis and interpretation. This study addresses current research challenges through a detailed comparative analysis of convolutional neural network (CNN) architectures, including pre-trained models like Inception V3, Xception, DenseNet-121, VGG-19, and a Custom CNN developed from scratch. The evaluation focuses on their effectiveness in detecting COVID-19 using a relatively vast and comprehensive dataset. The evaluation focuses on assessing the efficacy of these models in detecting COVID-19 across binary-class and multi-class classifications within a large dataset, offering valuable insights into their overall effectiveness. Grad-CAM is employed in this study, which visualizes picture salient areas affecting deep neural network predictions and improves model interpretability. The heatmaps reveal the model's decision-making processes, making their behavior more understandable to medical practitioners. With resulting accuracies of 99.13% and 97.66% in binary-class and multi-class classification, Inception V3 demonstrated the most optimal performance. This proposed approach surpasses existing benchmark methods, underscoring its superiority.