CT-CXR-Net: Optimal Deep Learning Framework for Dual-Modal COVID-19 Classification

Since the onset of the COVID-19 pandemic, more than 700 million people have been impacted by the disease, and more than 7 million people have died, underlining the necessity of quickly and efficiently identifying and diagnosing the disease in controlling its spread. Despite the immense progress, conventional methods of testing usually experience the drawback of being too slow, accessible, and imprecise, especially within resource-limited settings. Existing COVID-19 classification models based on Artificial Intelligence (AI) are affected by noise interference of the medical images, sub-optimal segmentation, and inefficient feature selection, all of which contribute to a low reliability of diagnosis. To overcome these challenges, the novel CT-CXR-COVID-19 Classification Network (CT-CXR-Net) method begins with Block-Matching and 3D filtering (BM3D) denoising to effectively eliminate complex noise patterns, ensuring high-quality input for further analysis. An Optimal U-Net (OU-Net) segmentation model is employed, whose loss is minimized using Modified Grey Wolf Optimization (MGWO), leading to precise lung region extraction. Subsequently, ResNet50 is utilized for deep feature extraction, capturing complex and informative patterns from both Computer Tomography (CT) and Chest X Ray (CXR) images. To reduce feature dimensionality and enhance classification performance, Improved Brown Bear Optimization (IBBO) is adopted for optimal feature selection. Finally, a Ridge Classifier provides robust and efficient classification, maintaining a balance between bias and variance. This approach achieves exceptional results on separate datasets, with the CT dataset recording 100% accuracy, while the CXR dataset achieves 99.30% accuracy, 99.60% precision, and 99.95% recall and F1-score, demonstrating its potential for reliable and high-performance COVID-19 diagnosis.