COVID-19, first appearing in 2019, rapidly spread globally, causing severe pneumonia and significant mortality. While vaccines exist, early and accurate diagnosis remains crucial for effective patient management. Manual diagnosis methods are limited, highlighting the need for automated systems. Artificial intelligence (AI) and machine learning (ML), particularly Artificial Neural Networks (ANNs), offer a potential solution. ANNs have shown promise in image segmentation and classification tasks, making them suitable for analyzing medical images like chest X-rays and CT scans. Multi-Layer Perceptrons (MLPs), a type of ANN, have been successfully employed in diagnosing various diseases. Optimizing MLP parameters, such as input features, hidden node size, initial weights, and learning rate, is crucial for achieving high diagnostic accuracy. Swarm intelligence algorithms, like the Artificial Bee Colony (ABC) algorithm, are effective for fine-tuning these parameters. This paper introduces a novel CAD system, ABCNN, leveraging ABC optimization to improve the accuracy and efficiency of COVID-19 detection from lung CT images. The system combines image preprocessing, region of interest (ROI) extraction, feature extraction, feature selection, and classification stages, all optimized using the ABC algorithm. The novelty of this approach lies in the simultaneous optimization of the segmentation, feature selection, and classification processes using ABC.

Several studies have explored the use of AI and ML for COVID-19 diagnosis using chest X-rays and CT scans. Some employed deep learning models like Convolutional Neural Networks (CNNs) and combined them with other techniques such as XGBoost for classification. These methods often focused on feature extraction from CT images followed by classification, achieving varying degrees of accuracy. However, many of these approaches required large datasets for training, and some relied on pre-trained models. Other research explored different optimization strategies to improve the performance of deep learning models for COVID-19 detection, including adaptive feature selection and voting classifier algorithms. Some studies used ensemble techniques combining multiple deep learning models to enhance accuracy. While these studies demonstrated the potential of AI in COVID-19 diagnosis, there's still a need for improved methods that can achieve high accuracy with smaller datasets and less computational complexity. The proposed ABCNN approach aims to address these challenges.

The proposed ABCNN system consists of several key steps: 1. **Image Acquisition:** The system uses a publicly available dataset of 470 lung CT images (275 COVID-19, 195 non-COVID-19) from [67]. The dataset is split into training (50%), validation (25%), and testing (25%) sets. 2. **Preprocessing:** A Wiener filter is applied to reduce noise while preserving image details. This improves the quality of the input images for subsequent processing steps. 3. **Region of Interest (ROI) Extraction:** This two-stage process first extracts lung regions using histogram-based thresholding. Then, an ABC-optimized region growing method segments infected areas within the extracted lung regions. The ABC algorithm determines the optimal threshold for the region growing process, improving segmentation accuracy. 4. **Feature Extraction:** Texture features (using Gray-Level Co-occurrence Matrix (GLCM) and Gray-Level Run-Length Matrix (GLRLM)) and intensity features (histogram-based) are extracted from the segmented infected regions. 5. **Feature Selection and Classification:** An ABC-optimized ANN classifier is used. The ABC algorithm simultaneously optimizes the input feature subset, initial weights, and the number of hidden nodes in the ANN. The ANN is trained using resilient backpropagation (RP), Levenberg-Marquardt (LM), and gradient descent (GD) algorithms and the performance is evaluated based on validation set error. The training stops when the validation error increases for six consecutive iterations. The best performing ANN configuration (in terms of accuracy and complexity) is selected for testing. The ABC algorithm used for optimization works by iteratively improving solutions representing possible threshold values for region growing, feature subsets, ANN weights, and the number of hidden nodes. This is done through the actions of employed bees, onlooker bees, and scout bees, inspired by the foraging behavior of honey bees. The fitness of each solution is evaluated based on the performance of the resulting ANN on the validation set. The best solution from each run is saved. Finally, the chosen ANN model is tested using the testing dataset.

The proposed ABCNN approach, using resilient backpropagation (RP), achieved a mean classification accuracy of 92.37% with a generation size of 20 in ten independent runs on the test dataset. This accuracy is significantly higher than the results obtained using the Levenberg-Marquardt (LM) and gradient descent (GD) backpropagation algorithms. Furthermore, the ABCNN-RP approach resulted in a less complex ANN network compared to the other backpropagation methods. The mean number of connections in the network using ABCNN-RP is 17.25, which is 9.35% less than ABCNN-GD and 14.65% less than ABCNN-LM. The ABCNN-RP also showed high sensitivity and specificity values indicating its effectiveness in correctly identifying both COVID-19 and non-COVID-19 cases. Moreover, the ABCNN-RP with feature selection exhibited superior classification accuracy and lower complexity compared to the ABCNN-RP without feature selection, demonstrating the effectiveness of the proposed feature selection method. Comparing ABCNN-RP to other swarm intelligence algorithms (Bat Algorithm, Bacterial Foraging Optimization, Ant Colony Optimization, Dragonfly Algorithm, and Particle Swarm Optimization) on the same dataset showed that ABCNN-RP achieved higher accuracy. Finally, the ABCNN method also outperformed other classifiers (SVM, Naive Bayes, Random Forest, and Decision Tree) when applied to the same dataset. These results demonstrate the superiority of the proposed ABCNN approach for COVID-19 detection from lung CT images.

The high accuracy and low complexity of the ABCNN-RP approach demonstrate its potential for use in a clinical setting. The simultaneous optimization of segmentation, feature selection, and ANN parameters using the ABC algorithm proves beneficial. The improved performance compared to other swarm intelligence algorithms and traditional classifiers highlights the effectiveness of the proposed method. The superior results compared to existing works demonstrate the significant contribution of this research to the field of COVID-19 diagnosis using AI. The findings support the use of ABCNN for assisting in the early and accurate detection of COVID-19, which is crucial for timely intervention and improved patient outcomes. Future research could explore the application of the ABCNN framework to larger and more diverse datasets, potentially incorporating deep learning models and cloud computing for enhanced scalability and accessibility in e-Healthcare applications.

This paper presents ABCNN, a novel CAD system for COVID-19 detection using lung CT images. The system leverages the ABC algorithm for simultaneous optimization of image segmentation, feature selection, and ANN parameters, leading to high accuracy and low complexity. ABCNN-RP outperforms other approaches, demonstrating its potential for real-world applications. Future work could investigate the application of this framework to larger datasets and integration with cloud-based e-Healthcare systems.

While the ABCNN approach demonstrates high accuracy, the study's generalizability might be limited by the use of a single publicly available dataset. The computational cost of the ABC optimization process might also be relatively high compared to simpler methods, though this is offset by the improved accuracy. Further research using different and more extensive datasets is required to validate these findings and assess the robustness of the approach across various imaging conditions and patient populations.