COVID-19 CLASSIFICATION IN CT IMAGES WITH CONVOLUTIONAL NEURAL NETWORK-BASED ENSEMBLE LEARNING
Loading...
Date
2022-04
Authors
Kushenchirekova, Dina
Journal Title
Journal ISSN
Volume Title
Publisher
Nazarbayev University School of Engineering and Digital Sciences
Abstract
The coronavirus infection has spread all over the world with great speed and the
virus continues to grow and change. The COVID-19 infection that became a cause
of the pandemic was a huge issue that people faced. Deep learning has a significant
and important part in application of medical image analysis, and in this paper we use
deep learning and convolutional neural network (CNN) methods. CNN helps us to
classify our formations, since it is an effective tool at image classification. Deep learning
is the field of Artificial Intelligence that copes with the classification problems,
such as classifying and recognizing COVID-19 infection using computer tomography
(CT) images that contain lungs. In the study, we utilize several of the most popular
convolutional neural networks and evaluate them using the common metrics. Among
8 CNN architectures we used, which are VGG-19, VGG-16, MobileNetV2, Xception,
ResNet50V2, DenseNet201, Inception-V3, and EfficientNetB3, the most efficient and
outperforming was VGG-19, as it achieved the highest accuracy score. Specifically,
the VGG-16 CNN architecture’s accuracy on CovidX CT dataset is 0.97, on SARSCoV-
2 CT dataset is 0.95, and on UCSD COVID-CT dataset the score is 0.94. The
arisen question now is how to properly utilize data mining to build an efficient detection
system and mining framework. To answer the question we decide to use ensemble
learning, which integrates fusion, modeling, and mining into a single model. Our proposal
is ensemble learning algorithm that substantially stacks several neural network
architectures into one. The logic behind the method is to extract features from the
images using several of the above-mentioned models and combine the features into a
"stack". The results suggest that the method performs better than each individual
architecture. As the ensemble model considers each of the features and the losses
provided by the models, the resultant loss is lower. This results in a higher accuracy
score. In this way, we achieved the Ensemble model’s accuracy of 0.9867 for the
UCSD COVID-CT dataset, while the highest accuracy of the individual model was
0.945. As a result of the SVM integrated alternative methodology, ensemble model
has shown the accuracy of 0.982 for SARS-COV-2 CT dataset.
Description
Keywords
SARS-COV-2 CT dataset, Type of access: Open Access, COVID-19, CNN, deep learning, convolutional neural networks, Research Subject Categories::TECHNOLOGY
Citation
Kushenchirekova, D. (2022). COVID-19 CLASSIFICATION IN CT IMAGES WITH CONVOLUTIONAL NEURAL NETWORK-BASED ENSEMBLE LEARNING (Unpublished master's thesis). Nazarbayev University, Nur-Sultan, Kazakhstan