COVID-19 DETECTION FROM X-RAY IMAGES USING DEEP LEARNING

Abstract

Currently, the detection of coronavirus is one of the main challenges in the world. Recent statistics have shown that the total number of cases are increasing exponentially. Existing high-precision diagnostic technologies such as RT-PCRs are expensive and complex. In order to obtain a quick and precise medical diagnosis, X-ray images are commonly used. Detecting positive cases of COVID 19 from X-ray images is really difficult, challenging and susceptible to human error. Various deep learning networks have been used in recent studies for X-ray image classification and have generated competitive results, because stages like feature selection, feature extraction and classification, are performed automatically in deep learning techniques. Coronavirus detection using various chest xray image data sets is a daunting task. Researchers have used a variety of preprocessing techniques, feature extraction methods and classification models. It is hard to suggest a method or a combination of methods that yield best results in detecting COVID-19 from xray images. In most articles, more than 90% accuracy was reported, which could be considered really high. However, the aim would be to increase the level of accuracy to almost 100%, as incorrect classification of the disease, even in a few cases, is totally unacceptable. In this project, a set of seven CNN based models (VGG16, VGG19, InceptionV3, Resnet50, Xception, Densenet121 and InceptionResNetV2) have been implemented for the detection of coronavirus infection using an open source dataset of x-ray images. A custom data set of 1000 covid positive and 1000 normal case xray images was generated for the experiment. The pretrained ImageNet weights were used, and a custom fully connected layer head with 5 layers was implemented for training the model for the covid dataset. Densenet121 outperformed other models with an accuracy of 97%.