The outbreak of COVID-19 has been a global crisis, impacting most countries and territories around
the world. Despite the social distancing and restrictive movement policies, cases have significantly
grown during the past three years. During the beginning of the outbreak, the cause of and treatments
for the corona virus were unclear. However, over time similar symptoms across various patients have
occurred, such as shortness of breath, headaches, persistent pain or pressure in the chest, fever,
nausea, and fatigue. More detailed observations of these symptoms with X-rays or CT scans show
patchy and confluent opacity and consolidation distributed throughout the lung zone developed over
time. Chest X-rays and CT scans can be a critical warning and useful diagnosis of COVID-19 and
COVID-19 pneumonia, also used during in the follow-up procedure of recovery [5]. Thus, we believe
a technological approach to such images can improve accuracy and speed of diagnosis and recovery,
showing an end to the long tunnel of the pandemic.
Midterm Report
Introduction
Related Work
A number of works present works of analyzing chest X-rays to detect COVID-19, mostly through
Deep Learning models such as Convolutional Neural Networks (CNN) [1, 3, 2, 7, 4]. Moreover, the
effectiveness of different architectures within CNN model have been evaluated, such as AlexNet, VGG16, GoogleNet, MobileNet-V2, SqueezeNet, ResNet-34, ResNet-50 and Inception-V3 for classifying
COVID-19 cases from normal cases. In related work, hyperparameters have been tuned within these
models including batch size, learning rate, number of epochs, and type of optimizers through a
comparative analysis to achieve the best suited model. However, most of prior work has been focused
on CNN models. In this project, we aim to investigate different possibilities of models including
XGBoost classifier, multinomial logistic regression, anomaly detection and support vector machines.
These approaches are often used for image classification, and we aim to understand why or why not
different approaches are effective.
Methodology
We plan to use the dataset from Kaggle (https://www.kaggle.com/datasets/bachrr/covid-chest-xray) which includes COVID-19 cases with chest X-ray or CT
images(see Figure a). It contains COVID-19 cases as well as MERS, SARS, and ARDS. Labels are
arranged in hiearchy in Figure b.
(a) Chest X-ray of COVID
(b) Labels in Hierarchy
As we plan to use various models we might need to preprocess the images and resize them to be fed
into the models. A technique called ”lung segmentation” has also been proposed in previous research
[6] where it segments the lung and discards rest of the CXRs to prune away potential bias such as the
presence of a medical device. While training we need to consider that the COVID datasets are built
mainly of positive cases therefore the models will be trained with a k-fold cross-validation procedure
which involves splitting the training dataset into k folds. However this might not be enough for
handling imbalanced datasets so tuning parameters will be considered such as balancing the positive
and negative weights during the training process.
Progress
For the classification baseline, we use simple logistic regression. We split the training and test
dataset by 0.75 and 0.25. The data sets we used are from Kaggle (https://www.kaggle.com/tawsifurrahman/covid19-radiography-database, https://www.kaggle.com/datasets/pranavraikokte/covid19-image-dataset).
For the CNN and XGBoost, we used a dataset from Kaggle (https://www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia) but from a different source. The
CNN model is built with 4 layers including dropout layers to avoid overfitting. We also use the Adam
optimizer for stochastic gradient descent and binary cross entropy for the loss function. XGBoost is
an scalable, flexible, and distributed machine learning library which provides parallel tree boosting
for better regression, classification, and ranking in learning problems. Our XGBoost model inputs the
output from the CNN model which is the probability of COVID-19 infection based on the patient’s
chest X-ray, age, and gender.
Future Work
Evaluation
Originally we planed to use a dataset from https://www.kaggle.com/bachrr/covid-chest-xray. However now we are testing multiple datasets from Kaggle to improve the performance of the modelsby combining and preprocessing data images. We used the validation set approach by splitting thedata to training/testing each containing Covid, Viral Pneumonia, and Normal Chest X-Rays.
For logistic regression, we assessed the quality of K-Means clusters (n = 3) and the Sillouette
Score was 0.156, which indicates that there are some overlapping clusters. Also, despite that the
XGBoost model provides more precise output by integrating CNN as a trainable feature extractor to
automatically obtain features from the input it’s accuracy was lower than CNN. Therefore we still
need to tune the parameters to obtain an optimistic result.
Timeline
We plan to implement other models such as few shot learning with siamese networks due to small
amount of data and finish developing the website for the presentation.
References
[1] R. Jain, M. Gupta, S. Taneja, and D. J. Hemanth. Deep learning based detection and analysis of
covid-19 on chest x-ray images. Applied Intelligence, 51(3):1690–1700, 2021.
[2] A. I. Khan, J. L. Shah, and M. M. Bhat. Coronet: A deep neural network for detection and
diagnosis of covid-19 from chest x-ray images. Computer methods and programs in biomedicine,
196:105581, 2020.
[3] S. R. Nayak, D. R. Nayak, U. Sinha, V. Arora, and R. B. Pachori. Application of deep learning techniques for detection of covid-19 cases using chest x-ray images: A comprehensive study.
Biomedical Signal Processing and Control, 64:102365, 2021.
[4] C. Ouchicha, O. Ammor, and M. Meknassi. Cvdnet: A novel deep learning architecture for detection of coronavirus (covid-19) from chest x-ray images. Chaos, Solitons & Fractals, 140:110245,
2020.
[5] L. A. Rousan, E. Elobeid, M. Karrar, and Y. Khader. Chest x-ray findings and temporal lung
changes in patients with covid-19 pneumonia. BMC Pulmonary Medicine, 20(1):1–9, 2020.
[6] E. Tartaglione, C. A. Barbano, C. Berzovini, M. Calandri, and M. Grangetto. Unveiling covid-19
from chest x-ray with deep learning: A hurdles race with small data. International Journal of
Environmental Research and Public Health, 17(18):6933, Sep 2020. URL: http://dx.doi.org/
10.3390/ijerph17186933, doi:10.3390/ijerph17186933.
[7] L. Wang, Z. Q. Lin, and A. Wong. Covid-net: A tailored deep convolutional neural network design
for detection of covid-19 cases from chest x-ray images. Scientific Reports, 10(1):1–12, 2020.