Validation of COVID-19 Grading System based on Harris Hawks Optimization (HHO) and Variational Quantum Classifier using JLDS-2024

Dr.Javeria Amin¹nadia gul²Muhammad Sharif³Steven Fernandes^4*

• ¹University of Wah, Wah Cantonment, Pakistan
• ²Wah Medical College, Wahcantt, Pakistan
• ³COMSATS University Islamabad, Wah Campus, Wah Cantt, Pakistan
• ⁴Department of Journalism, Media & Computing, College of Arts and Sciences, Creighton University, Omaha, United States

Introduction: COVID-19 is an infectious respiratory disease that is caused by the SARS-CoV-2 virus. It is mainly transmitted by respiratory droplets that may produce mild symptoms to severe sickness or death. COVID-19 attacks both the trachea and lungs. High-resolution computed tomography (HRCT) scans are preferred for diagnosing several respiratory disorders at an early stage and more accurately.Accurate detection of COVID-19 using CT images is still challenging due to several factors, such as low-quality images, illumination, and variation in the shape and size of the COVID-19 lesions. Therefore, the main objective is to design a deep learning model for accurate detection at an early stage.Method: A method is proposed for classifying and segmenting COVID-19 lesions. First, preprocessing is performed on axial, coronial, and sagittal views of CT slices using statistical methods based on volumetric analysis to compute the scores. These are used to assign the grades based on the severity level of each lung lobe such as mild, moderate, and healthy. The graded slices are input to the proposed classification model, and N × 1000 features are extracted at the fully connected (FC-1000) layer of the pre-trained ResNet-18 model. Out of these N × 469 features are selected through the Harris Hawks Optimization (HHO) model. Finally, classification is performed using variational quantum, neural network, and Naïve Bayes classifiers. To segment the classified images, fine-tuned U-net model is designed having 47 layers such as 1 input, 5 dropout, 11 conv, 9 ReLU, 9 Batch-normalization, 4 Conv2DTranspo, 4 Concatenate, and 4 max-pooling. The model is trained from scratch based on optimal hyperparameters such as binary cross-entropy loss, 16 filter size, 0.005 drop-out, 8 batch size, and Adam optimizer.The proposed method is evaluated on local and publicly available UCSD-AI4H/COVID-CT datasets and provides a mean-ROC of 0.98 ± 0.01 that strengthen the novelty of the proposed method.