A convolutional neural network based approach for Covid 19 classification

Abstract

COVID-19, declared a global pandemic by WHO, spread exponentially and created severe health, economic, and social disruption worldwide. Early detection and isolation are essential for controlling transmission. Although RT-PCR is the most common diagnostic method, it suffers from limitations such as high processing time, low sensitivity, false negatives, and variability in testing procedures. CT-scan imaging can detect COVID-19 even in asymptomatic or RT-PCR-negative patients, but manual diagnosis is challenging due to workload and human error. This thesis addresses these challenges by proposing AI-based automated techniques for classifying COVID-19 using CT-scan images. First, it presents a comprehensive survey of existing COVID-19 classification techniques, COVID-19 datasets, and their related challenges. Second, it proposes a Deep Learning-Based Hybrid Approach (DLBHA) that uses Contrast Limited Adaptive Histogram Equalization (CLAHE) for image enhancement and Atom Search Optimization for optimizing DCNN hyperparameters. Third, it introduces the Social Distancing-Induced Coronavirus Optimization (COVO) algorithm, a novel metaheuristic validated on 13 benchmark functions and real-world IEEE-CEC 2011 problems, showing superior convergence and accuracy. Fourth, a deep CNN classifier optimized with Intelligent Wolf Optimization is developed, utilizing texture features extracted from lung regions through LOOP, LBP, and ResNet-101 descriptors. Fifth, a robust COVID-19 detection framework is developed, integrating LDP-RCM and GLCM for enhanced feature extraction, along with IDCNN-COVO for improved classification accuracy. newline