Early and accurate detection of COVID-19 is crucial. While PCR testing is a standard method, chest X-rays and CT scans offer non-contact alternatives for rapid assessment. However, manual interpretation of these images is time-consuming and requires expertise. Deep learning (DL) techniques provide a potential solution for automating this process. This research addresses the challenges of automatically classifying COVID-19 and pneumonia from medical images by leveraging the power of DL and exploring various architectures and transfer learning strategies. The goal is to develop a model that achieves high accuracy, precision, and recall, surpassing the performance of existing methods, thus facilitating faster and more efficient diagnosis.

Existing literature highlights the application of various machine learning (ML) and deep learning (DL) techniques for COVID-19 detection using chest X-rays and CT scans. Traditional methods like Support Vector Machines (SVMs) and backpropagation networks have been used, but DL models, particularly Convolutional Neural Networks (CNNs), have shown superior performance. Several studies explored pre-trained CNN architectures like ResNet, DenseNet, and VGG-16 through transfer learning, adapting them for COVID-19 detection. These models have demonstrated promising results but often suffer from limitations such as overfitting and a dependence on large, well-annotated datasets. This research builds upon these efforts, aiming to address these limitations and improve diagnostic accuracy by exploring ensemble learning and advanced image pre-processing techniques.

The proposed model comprises several key stages. Initially, spatial domain filtering (mean, Wiener, and median filters) is applied to enhance the quality of input images by reducing noise and improving contrast. This is followed by histogram equalization to further enhance image contrast and improve feature extraction. Feature extraction and classification are performed using two primary deep learning architectures: 1. **Fast Attention-Based ResNet:** This architecture incorporates an attention module to improve feature extraction by weighting features based on their relevance. The attention module calculates the relationship between features to highlight the most relevant information. The ResNet model then performs classification using a softmax layer. The learning rate is set to 0.0001, and training is performed for 10 epochs. 2. **Enhanced VGG-16:** This architecture is specifically designed for X-ray image classification. It uses a modified VGG-16 structure with three layers: convolutional, fully connected, and softmax. Max pooling and average pooling layers are used for feature extraction, and these results are concatenated using a sigmoid function. This model is optimized for multi-class classification. The proposed model uses an ensemble approach by combining the predictions from both the ResNet and enhanced VGG-16 architectures to obtain final classification results. Data augmentation techniques are employed to augment the training datasets, increasing the diversity of training samples and mitigating overfitting issues. Five publicly available datasets are used for evaluation, two containing X-ray images and three containing CT scan images.

The proposed ensemble model exhibits significantly improved performance compared to existing methods. The model achieves a high classification accuracy of 95–96% and a precision of 95–96%. The recall (sensitivity) is also high, ranging from 94% to 96%, indicating a low rate of false negatives. The F-measure, which combines precision and recall, reaches 95–97%, demonstrating a strong overall performance. The computational time for the proposed model is relatively low, less than 0.5 seconds, showing its efficiency. The superior performance is demonstrated through comparison with various other models including CNNs, DNNs, and transfer learning based models. Detailed comparisons across multiple performance metrics (accuracy, precision, recall, F-measure) are shown in tables and figures, showcasing the significant improvements achieved by the proposed ensemble approach. The results are consistent across different datasets and iteration counts, underscoring the robustness of the proposed methodology.

The superior performance of the proposed ensemble model can be attributed to the combination of effective image pre-processing, attention-based feature extraction, and the ensemble strategy. The spatial domain filtering and histogram equalization effectively reduce noise and improve contrast, leading to better feature representation. The attention mechanism in the ResNet model enhances the ability to focus on the most informative features, while the enhanced VGG-16 model provides a complementary approach tailored for X-ray images. The ensemble approach combines the strengths of both models, yielding a more robust and accurate classification system. These findings demonstrate the potential of the proposed method for improving the speed and accuracy of COVID-19 and pneumonia diagnosis.

This research successfully developed a deep learning ensemble framework for the accurate and efficient detection of COVID-19 and pneumonia from CT scans and X-ray images. The model demonstrates superior performance compared to existing methods, achieving high accuracy, precision, recall, and F-measure. Its low computational time makes it suitable for practical implementation in clinical settings. Future research could explore the integration of other deep learning architectures, the use of larger and more diverse datasets, and the incorporation of explainable AI techniques to enhance the interpretability of the model's predictions.

The performance of the model might be affected by the quality of the input images. Noisy or poorly acquired images could potentially reduce the accuracy of the model. The dataset used for training and evaluation, while large, may not fully represent the diversity of COVID-19 and pneumonia cases across various populations and imaging techniques. The generalizability of the model to other unseen datasets needs further evaluation. Future studies should investigate these limitations and explore methods to mitigate them.