Hybrid COVID-19 segmentation and recognition framework (HMB-HCF) using deep learning and genetic algorithms

Abstract

COVID-19 (Coronavirus) went through a rapid escalation until it became a pandemic disease. The normal and manual medical infection discovery may take few days and therefore computer science engineers can share in the development of the automatic diagnosis for fast detection of that disease. The study suggests a hybrid COVID-19 framework (named HMB-HCF) based on deep learning (DL), genetic algorithm (GA), weighted sum (WS), and majority voting principles in nine phases. Its segmentation phase suggests a lung segmentation algorithm using X-Ray images (named HMB-LSAXI) for extracting lungs. Its classification phase is built from a hybrid convolutional neural network (CNN) architecture using an abstractly-designed CNN (named HMB1-COVID19) and transfer learning (TL) pre-trained models (VGG16, VGG19, ResNet50, ResNet101, Xception, DenseNet121, DenseNet169, MobileNet, and MobileNetV2). The hybrid CNN architecture is used for learning, classification, and parameters optimization while GA is used to optimize the hyperparameters. This hybrid working mechanism is combined in an overall algorithm named HMB-DLGA. The study experiments implemented the WS approach to evaluate the models' performance using the loss, accuracy, F1-score, precision, recall, and area under curve (AUC) metrics with different pre-defined ratios. A collected, combined, and unified X-Ray dataset from 8 different public datasets was used alongside the regularization, dropout, and data augmentation techniques to limit the overall overfitting. The applied experiments reported state-of-the-art metrics. VGG16 reported 100% WS metric (i.e., 0.0097, 99.78%, 0.9984, 99.89%, 99.78%, and 0.9996 for the loss, accuracy, F1, precision, recall, and AUC respectively) concerning the highest WS. It also reported a 99.92% WS metric (i.e., 0.0099, 99.84%, 0.9984, 99.84%, 99.84%, and 0.9996 for the loss, accuracy, F1, precision, recall, and AUC respectively) concerning the last reported WS result. HMB-HCF was validated on 13 different public datasets to verify its generalization. The best-achieved metrics were compared with 13 related studies. These extensive experiments' target was the applicability verification and generalization.