IoT-Based Classification and Differential Diagnosis of COVID-19 from CT Images using DexiNed Filter, Image Blending and Weighted Ensemble CNNs

Abstract

The coronavirus, discovered for the first time in December 2019 in Wuhan, China, quickly spread to more than two hundred countries and became a public health emergency. After more than three years, the disease is making a comeback. This article aims to improve the automated detection of COVID- 19 infection in CT images by introducing a deep learning-based approach within the context of the Internet of Things (IoT). First, with the assistance of a specialist physician, we collected data in four classes: normal, COVID-19, viral pneumonia, and bacterial pneumonia, and categorized them as COVID and non-COVID. Then, inspired by the image blending, we merged two features of classes, such as COVID-19 and viral pneumonia, to create a new synthetic COVID-19 image. Even though such images occurs infrequently, they plays a crucial role in differential diagnosis. With this idea, the problem of how to collect them will be solved. The DexiNed Alter was used to extract deep features from these images, which were then used as inputs to the convolutional neural networks AlexNet, ResNet50, DenseNet201, VGG16, and InceptionV3 for classification. The weighted ensemble of these models achieved an accuracy of 95.34%, 95% precision, 95.54% sensitivity, 95.15% specificity and 95.27% F1-score.

Ähnliche Arbeiten