COVID-19 Prognosis from Chest X-ray Images by using Deep Learning Approaches: A Next Generation Diagnostic Tool

The study addresses the need for rapid, automatic screening of COVID-19 amid recurrent SARS-CoV-2 variants (e.g., Delta, Omicron and sub-lineages) that have strained global healthcare systems. Shortages of tests and medical resources, alongside ongoing transmission and multiorgan impacts, underscore the importance of AI-based tools for early detection. Prior ML/DL approaches have been investigated for COVID-19 classification from CXR images. The objective here is to classify CXR images into normal, pneumonia, and COVID-19 using VGG16 and LSTM models to support early prognosis and clinical decision-making.

The paper cites several AI/DL studies for COVID-19 detection from CXR: a triple-view CNN leveraging left/right/overall lung views for binary healthy vs COVID-19 classification (Zhang); a custom CNN achieving 98.3% accuracy but unable to distinguish COVID-19 from other pneumonias (Haque and Abdelgawad); Bayesian CNN with VGG16 features reporting 85.2–92.9% accuracy (Ghoshal and Tucker); a combined CNN-LSTM model achieving 99.4% accuracy with high sensitivity/specificity (Islam et al.); InceptionV3, ResNet50, InceptionResNetV2 models with up to 98% accuracy on small binary datasets (Narin et al.); and a DL fusion framework via transfer learning achieving 95.49% accuracy (Shorfuzzaman et al.). A comparative table lists accuracies of various architectures (e.g., VGG-19, ResNet-50, DenseNet) and reports the proposed VGG16 at 97% and LSTM at 92%.

Dataset: Public COVID-19 CXR dataset from https://github.com/ieee8023/covid-chestxray-dataset. The text reports differing counts: one section notes 5,863 images (94 normal, 506 COVID-19, 46 pneumonia), while another states 6,432 images total with 20% used for testing. Images were curated to remove low-quality/unreadable cases. Data were split into train/test (80/20). Implementation in Python (TensorFlow/Keras) with Pandas, NumPy, matplotlib. Training settings: 200 epochs, batch size 32, learning rate 0.0001, cross-entropy loss, ADAM optimizer (referred to as loss function in text), SoftMax classifier for three classes (normal, COVID-19, pneumonia). Models: VGG16 CNN architecture with convolution, pooling, two fully connected layers, and a softmax output; reported total parameters ~134,272,835 (trainable all). LSTM (RNN) architecture with input, forget, and output gates; equations for gate operations provided; used for sequence modeling to capture dependencies in extracted features. Evaluation: Confusion matrices, ROC curves, precision, recall, F1-score, training/validation accuracy and loss were computed for both models.

- ROC: VGG16 achieved ROC area of 100% for COVID-19 and 99% for pneumonia; LSTM achieved 97% (COVID-19), 98% (pneumonia), and 96% (normal). - Precision/Recall/F1 (as described in text section): VGG16 — COVID-19: 0.95/0.95/0.95; Normal: 0.91/0.95/0.93; Pneumonia: 0.95/0.90/0.93. LSTM — COVID-19: 0.86/0.82/0.84; Normal: 0.95/0.86/0.90; Pneumonia: 0.79/0.90/0.84. - Precision/Recall/F1 (as per Table 3, which differs from narrative): VGG16 — Normal: 0.96/0.96/0.96; COVID-19: 0.90/0.88/0.89; Pneumonia: 0.95/0.96/0.96. LSTM — Normal: 0.99/0.96/0.97; COVID-19: 0.89/0.91/0.90; Pneumonia: 0.96/0.96/0.96. - Training/Validation accuracy and loss: LSTM reached 92% validation and 88% training accuracy after 200 epochs; validation loss 0.01 and training loss 0.001 at 11 epochs; validation accuracy peaked around 35 epochs. VGG16 showed better overall performance than LSTM; elsewhere in the paper's conclusion it reports LSTM 98% validation and 92% training accuracy, and VGG16 100% validation and 98.94% training accuracy, indicating internal reporting inconsistencies. - Comparative analysis suggests the proposed VGG16 model achieved ~97–97.5% accuracy, outperforming several listed prior methods, though with higher parameter count (~134M).

The findings indicate that deep learning models can effectively classify CXR images into normal, pneumonia, and COVID-19 categories, supporting early and automated screening. VGG16 generally outperformed LSTM across ROC, precision/recall/F1, and accuracy, suggesting CNN-based feature extraction is particularly effective for CXR image classification. High recall, especially for COVID-19 and pneumonia, is clinically valuable to minimize false negatives that could lead to missed cases. The results support the study’s goal of enabling rapid, computer-aided triage during resource-constrained scenarios such as pandemic peaks, potentially aiding clinicians by providing fast preliminary prognosis and prioritization. Comparative results against existing methods highlight competitive or superior performance, albeit with a larger model size. The authors note that integrating multi-source data (data fusion) and expanding datasets could further improve generalizability and robustness.

The study applied VGG16 and LSTM deep learning techniques to public CXR datasets for multi-class (normal, pneumonia, COVID-19) and binary classification. Reported performance shows VGG16 outperforming LSTM, with high validation and training accuracies and favorable ROC, precision, recall, and F1-scores. The authors state their models compare favorably to prior works, and emphasize that data fusion with larger datasets could further enhance diagnostic and predictive performance. The proposed approaches could assist clinicians with rapid prognosis of COVID-19 cases, supporting healthcare systems during surges. Future work will focus on multi-criteria classification to handle mixed lung infections (e.g., tuberculosis, AIDS, COVID-19).