The COVID-19 pandemic caused significant global health issues. Rapid and accurate diagnosis is crucial for effective management and prevention. While RT-PCR is a primary diagnostic method, it suffers from limitations like false positives and negatives. Computed Tomography (CT) scans offer higher accuracy but are expensive and expose patients to higher radiation. Chest X-rays (CXR) provide a cheaper and faster alternative. Several studies have shown the potential of deep learning models in diagnosing COVID-19 from CXR images. This study aims to analyze and evaluate the performance of five state-of-the-art deep learning models in identifying COVID-19 from CXR images, focusing on their individual capabilities and potential for real-world medical applications.

The introduction mentions several studies that explored using CT scans and deep learning methods for COVID-19 detection from CXR images. However, it notes some limitations, such as limited datasets used for training and evaluation. The authors highlight the need for a study using a larger dataset to assess the generalization capability of deep learning models in a clinical setting. The cited literature includes papers discussing RT-PCR limitations, CT scan advantages and disadvantages, and the use of AI in COVID-19 diagnosis.

The study utilized the COVID-QU dataset, comprising 33,920 CXR images categorized into COVID-19 positive, non-COVID-19 (viral or bacterial pneumonia), and normal. The dataset was divided into training, validation, and testing sets. Five pre-trained deep learning models (ResNet50, ResNet101, DenseNet121, DenseNet169, and InceptionV3) were used with transfer learning. Data augmentation (random rotation and horizontal flipping) was applied during training. Images were resized to 224x224 for ResNet and DenseNet models, while InceptionV3 used its native size. A custom classifier was added to each model, consisting of a global average pooling layer (or flatten layer for InceptionV3), a dropout layer, and a 3-unit dense layer with softmax activation. The Adam optimizer with categorical cross-entropy loss was used. Model training involved initial training with only the classifier and then fine-tuning by unfreezing layers. Evaluation metrics included categorical accuracy, precision, recall, and F1-score. Callbacks like model checkpoint, early stopping, and learning rate reduction were used. The programming language was Python with TensorFlow and Keras.

All five models achieved high recall values (>93%). ResNet101 exhibited the best overall performance, achieving 96% precision, recall, and accuracy. ResNet50 showed excellent precision, recall, and F1-score for the COVID-19 class (97%) but lower performance for non-COVID-19 and normal classes. ResNet101 also demonstrated high performance in COVID-19 classification (99% precision, 96% recall, 98% F1-score), maintaining balanced performance across all classes. DenseNet121 and DenseNet169 showed similar performance, with high precision for COVID-19 but lower recall for non-COVID-19 and normal. InceptionV3 reached 97% in precision, recall, and F1-score for COVID-19 but exhibited some misclassifications between normal and non-COVID-19. Confusion matrices are presented for each model, visualizing the classification results. The study highlights that ResNet101, despite having the largest number of trainable parameters, achieved the highest overall performance, suggesting that higher computational complexity may be beneficial.

The results demonstrate the potential of deep learning models for automated COVID-19 detection from CXR images. The high recall values are significant, as correctly identifying COVID-19 positive cases is crucial. ResNet101's superior performance could be attributed to its architecture's effectiveness in handling complex features in medical images. The use of a large dataset helped in achieving good generalization. Future work will involve incorporating lung segmentation and localization to improve accuracy, evaluating ensemble methods, and comparing the system's performance against radiologists to assess its clinical viability. The study acknowledges the limitations of focusing on individual models instead of ensemble techniques.

This study successfully evaluated five deep learning models for COVID-19 detection from CXR images. ResNet101 showed the best results, achieving 96% across all metrics. Future research should explore lung segmentation, ensemble models, and comparative testing with expert radiologists to enhance the system's accuracy and clinical applicability.

The study focuses on individual model performance rather than ensemble methods. The specific characteristics of the COVID-QU dataset and its potential biases might affect generalizability. Further validation with a broader range of CXR images from different sources and clinical settings is needed before clinical deployment. Comparison with radiologists' performance is also a crucial next step.