Fusion of Hand-Crafted and Automatically Generated features for improving the performance of COVID-19 X- ray image classification

ABSTRACT

In December 2019, the new disease COVID-19 was initially discovered in Wuhan, China and with a fast pace, it took over the whole world. It has impacted everyone's health, as well as the global economy and people's daily lives. It has become crucial that all the positive cases be detected quickly so it is possible to save other lives. Still, the lack of doctors and the lower availability of the test kits made it an arduous task. Recent research shows that radiological imaging techniques have played a valuable role in the detection of COVID-19. The use of artificial intelligence technology with radiological images can help identify the disease very accurately. Even in remote areas, it can be beneficial to overcome the shortage of doctors. This study proposed a method based on the aggregation of the extracted hand-crafted features with the automated ones. We used a his-togram of oriented gradients (HOG) for the hand-crafted features extraction. In addition, several techniques are investigated to get the deep learned features such as "DenseNet201","Inception ResnetV2", "VGG16","VGG19", "Inception_V3", "Resnet50", "MobileNet_V2"and "Xception"out of which "VGG19"gives optimal performance. Furthermore, for dimensionality reduction and to maintain the consistency of features, "principal component analysis (PCA)"is used. Our experiments on COVID-19 image datasets revealed that the proposed method achieves 99% classification accuracy in classifying normal and pneumonia X-ray images. © 2021 IEEE.