Detection of COVID-19 using deep learning on x-ray lung images

Introduction

The COVID-19 (WHO, 2019) virus is a fast-spreading disease that has appeared all over the world; Fig. 1 (Armstrong & Richter, 2021) represents the high spreading of COVID-19 in many countries as of February 8, 2021. It affects all aspects of life, economically, educationally, healthily, socially, and, most dangerously, human health, as it has caused the death of millions of people (Maatuk et al., 2022). There are two ways to control the spread of the virus: one is to take precautions, and the other is to take the vaccinations that are being specifically developed to defend against the disease. Studies have shown that vaccinations could cause allergic responses (Schumaker et al., 2020), thus some people may not be allowed to take it and would lead to the other way which is the early detection of the disease by chest x-ray to develop a treatment plan which provides a great opportunity for recovery and to prevent its spread.

Deep learning (Schulz & Behnke, 2012) has also proven to be useful in various fields including natural language processing (Vasilakes, Zhou & Zhang, 2020), computer vision (Voulodimos et al., 2018), healthcare (Esteva et al., 2019), and many other areas (Chatterjee, 2022). This research focuses on applying different deep learning approaches in the medical field. Many scholars have performed studies to help improve the medical field with computer-aid techniques to help doctors diagnose diseases and plan an early treatment process for the patients because sometimes tests can take a long time to receive results (Ayan & Ünver, 2019). One of these applications is using a convolutional neural networks (CNN) on medical images to extract features and train on them, for instance, using x-ray images (Chakraborty, Murali & Mitra, 2022) or CT-scans (Yadlapalli et al., 2022) to classify a new x-ray image as normal or diseased. As discussed earlier, diagnosing respiratory system diseases may be very useful for both doctors and patients. For this purpose, a dataset of radiography images of a patient’s chest is acquired to train different deep learning. The result of this training is a model that would be able to classify a new chest x-ray as normal or COVID-19. Some researchers have done studies on the diagnosis of COVID-19 (Padma & Kumari, 2020), lung opacity, and pneumonia infection (Kundu et al., 2021) in the x-rays while others have investigated all these diseases together (Khan et al., 2022).

Different approaches are provided in deep learning such as transfer learning. Many networks will be trained on huge datasets and have preserved the weights and parameters of those datasets. Transfer learning is defined as when using a pre-trained model or network for solving a specific task by freezing the earlier layers to preserve the pre-trained weights and adding some layers in the end for classification purposes and training some weights for the specific task (Bozinovski, 2020).

In this article, a dataset containing a set of lung X-ray images associated with corona disease was studied. Therefore, deep learning was chosen to train different neural networks and models including pre-trained ones on the images to determine whether the person was infected by COVID-19 or not and provide the needed assistance for them. Furthermore, the effects of manipulating and fine-tuning the hyperparameters of the networks are explored in addition to changing the architecture of the network to achieve satisfying and better results.

The rest of this article is organized as follows: the related work section discusses related work on deep learning and detecting diseases. The methodology section describes the proposed methods. The results section represents the results of the classification. The discussion section examines the results and interprets them. Finally, the conclusions and future work section to point out possible areas of research.

Related work

Deep learning has been used in different fields including the health sector. It is used to detect different kinds of diseases from x-rays, symptoms, or CT-Scans. Some of the research in this field is discussed and to be more specific, the research to detect respiratory diseases and COVID-19 is discussed.

For starters, the authors of Chakraborty, Murali & Mitra (2022), where another approach is presented for COVID-19 detection using deep learning. They stated that their system is for COVID-19 detection, but similar to Aggarwal et al. (2022) they classify the image into three classes. They used a larger dataset than the previous one and which contains 10,040 samples and has some imbalance issues. Still, they attempted to remedy the sample size issue through data augmentation as a data pre-processing step in addition to resizing the images to be fit for the model. The pre-trained models such as AlexNet, VGG, ResNet, and DenseNet were used for the comparison, but they have not mentioned any fine-tuning process for these networks in their work. The main strength of this article is using image segmentation to focus on the required area of the image which enhances the training quality and results of the model. The experiments were conducted using 70% training, 10% validation, and 20% testing along with 200 epochs and 50 batch sizes. The evaluation metrics that were taken into account were confusion matrix, accuracy, F1-score, precision, and recall in addition to the ROC curve. In the results section, the proposed model achieved a detection accuracy of 96.43% and a sensitivity of 93.68%. However, the sizes of the dataset even with data augmentation are not sufficient, it still needs more samples. Furthermore, none of hyperparameters comparison and nor fine-tuning of the networks were depicted.

Another research in this area was conducted by Yadlapalli et al. (2022) who studied the classification of CT scans for the diagnosis of COVID-19 patients. Their approach was to use K-means clustering as an image segmentation technique to separate the area of interest from the background and then feed the segmented images to the VGG16 network and another three-layer CNN built from scratch. The image segmentation led to smoother learning for both models even though it led to some kind of overfitting as it reduces the number of features in an image. A COVID-19-CT dataset was used that is distributed as 349 positive and 397 negative scans. For evaluation, they took into consideration the accuracy, F1-score, and area under the curve (AUC) for each model and their results showed that VGG16 achieved the highest results with an accuracy of 89%, F1-Score of 88%, and AUC OF 0.94 when compared with other networks.

Padma & Kumari (2020) built a CNN from scratch to identify COVID-19 patients. The motivation was to prevent further spread and initiate earlier treatment for existing cases. The dataset was acquired from open sources like GitHub and Kaggle and consists of 60 images where 30 of which are of COVID-19 patients and 30 of normal patients. The network used CNN layers, max-pooling to reduce dimensionalities, dropout to reduce the number of parameters, flattening to convert the image into a one-dimensional vector, and finally, a classification layer where the output is binary either positive or negative. ReLU was used as an activation function for the CNN layer which enhanced the performance by 2.5%. Furthermore, the dataset was split into 80:20 for training and testing. A training accuracy of 99.2% was acquired, with a validation accuracy of 98.8% and a loss of 0.3%. The precision of their study was 100% indicating that their model has no false-positive values.

Khan et al. (2022) compared the performance of three different pre-trained networks and used two strategies, where they added some regularization techniques to enhance the performance of COVID-19 detection. The dataset they used in their research comes from different sources such as Kaggle and GitHub and consists of 21,165 images distributed over four classes that are normal, lung opacity, viral pneumonia, and COVID-19. They mentioned that the dataset was imbalanced and used data augmentation to make it more balanced, but this only works to a certain extent, and collecting more data works better than this approach. They applied rotation, and horizontal and vertical flipping to produce more samples for their dataset in addition to rescaling and resizing the images as part of the pre-processing steps of their work. The split ratio of the dataset is 70:20:10. The comparison includes fine-tuning the hyperparameters of the networks (such as dropout percentage and learning rate) that are not clear. The results showed that EfficientNetB1 outperforms the other models, and that strategy II achieves better performance than strategy I. The recorded accuracy of the best performing model was 96.13% after choosing the confusion matrix, accuracy, F1-score, precision, and recall as the evaluation metrics.

Aggarwal et al. (2022) proposed a comparison between eight fine-tuned neural networks to classify x-ray images to different diseases represented as classes. Their work included working on two datasets where the first one consisted of three classes: normal, COVID-19, and pneumonia with a total of 709 samples, whereas the second dataset contains the same classes but focuses on the pneumonia class and divides it into bacterial and viral and consists of 959 samples. For experimenting, the datasets were split into 80% for training and 20% for testing. Because the sizes of the datasets are small, some data augmentation techniques are implemented to increase the sample size and make the model more generalizable. Furthermore, some preprocessing was done on the data such as applying CLAHE followed by image scaling and resizing. The metrics for evaluation that were used in this article include confusion matrix, accuracy, F1-score, precision, and recall. The acquired results showed that DenseNet121 achieved the best accuracy for the first dataset (97%) and the second dataset MobileNetV2 came (with 81%) with the best accuracy. The weaknesses and limitations of such a article are the sizes of the dataset even with data augmentation, it still needs more samples. In addition, the acquired results for bacterial and viral pneumonia were poor in terms of F1-score when compared with the results of the other classes and need further enchantments because samples of viral and bacterial pneumonia have been misclassified into other categories mainly into the COVID-19 category. Lastly, the fine-tuned architecture of the networks is not discussed and no experimenting with the hyperparameters is shown, which needs to be shown and discussed when doing comparative research.

Some researchers such as Nazish et al. (2021) trained their machine learning algorithms using x-ray images of the lungs for the classification into either COVID-19 or normal lungs. X-rays of normal lung and COIVID-19 patients were captured and connected to a dataset on Kaggle. They used these images and applied some pre-processing to them like removing noise and resizing. After that, they extracted features from the images using histogram oriented gradient (HOG). The dataset was split into 70% for training and the rest of 30% for testing. Two classifiers were applied to the extracted features which were SVM and logistic regression (LR). Their results have shown that the SVM has performed with better accuracy than LR with accuracies of 96% and 92% respectively.

From the literature that is discussed, the size of the dataset is the main issue that faces most researchers. In addition, most of these articles do not demonstrate a comparison of hyperparameters of the networks or how the pre-trained networks are fine-tuned for the specific intended task. Furthermore, a lot of the literature discusses respiratory diseases in general but a specialized model for detecting COVID-19 cases from x-ray images is needed instead of a model that can predict different diseases with less efficiency and accuracy. Some techniques that were used such as data augmentation, segmentation, and masking might come in handy and be used in our research to get better performance and results from the model.

Methods

Overview

The proposed model for detecting COVID-19 as shown in Fig. 2 contains three phases. Phase 1 demonstrates the data collection process and the manual pre-processing that is used to change image formats and sizes. Phase 2 represents sampling and splitting of the dataset into training and testing datasets followed by applying code pre-processing on the data to prepare them for Phase 3 where the models are trained on the training data then tested on the testing data and classification networks will be applied to classify the images into two categories (COVID-19, Normal). For further clarification of our approach, it is as follows: Data Collection (Kaggle); Data Preparation (resizing and renaming); Data Splitting; Data Preprocessing (depending on the network); Training; Evaluating (accuracy, loss, and confusion matrix).

Experiments and evaluation

After setting and implementing the architecture for the networks, it is time to run experiments on them. Many hyperparameters can affect the learning process of the network like the optimization function, activation function, learning rate, batch size, dropout percentage, and others. In our work, experiments are run with different values for learning rate, dropout percentage, batch sizes, and decay for learning rate. Table 4 shows the settings of the experiment. The first experiment included running each model with both learning rates and then choosing the one with better results. After that, another experiment with the decay values and freezing the learning rate to the one chosen earlier and comparing the results. The experiments for the dropout percentage and batch size are conducted the same way. The results and comparison are provided in the results and discussion sections. Adam optimizer is used as the optimization function for the network and early stopping is implemented with the patience of five along with 50 epochs it monitors the validation loss and returns the weights of the best epoch. For evaluating the developed models, several metrics are defined. The metrics that are used are the weighted accuracy, precision, recall, F1-score, confusion matrix, and learning curves for each model. Furthermore, the free version of Google Colab is used for conducting these experiments. It provides users with 13 GB of RAM. Tesla T4 P8 GPU, and 78 GB of storage with a maximum runtime of 6 hours that is shown in Fig. 3.

Table 4:

Expirement settings.

Parameter	Value
Learning rate	0.001/0.0001
Decay	Yes/No
Dropout percentage	0.5/0.2
Batch size	128/64/32

DOI: 10.7717/peerj-cs.1082/table-4

Results

In this section, the results of the experiments are presented and compared to leverage one of the networks to be the best-suited network for our dataset and use case. The comparisons are based on the training, validation, and testing loss. The network with the lowest testing loss and close to the validation and training losses is considered to be the best one as it is proof that the network does not overfit or underfit. Furthermore, in case of two networks have the same losses, comparisons are expanded to include reports such as classification report data and confusion matrix evaluation to determine the best one. This section is divided into two subsections which are the results of each transfer learning network where different comparisons on hyperparameters values are conducted to choose the best set of hyperparameters for each network. The second subsection compares the networks with the best set of hyperparameters of different types of networks to determine the best overall network. A detailed discussion of the results is provided in the following discussion section.

Comparisons of different networks

As shown above, the best-performing networks are compared to nominate the best network based on our dataset and experiments to be used in other applications. Table 10 summarizes the results of all the experiments and presents the best-performing network of each type.

Table 10:

Comparison of best performance models.

Model	Training accuracy (%)	Training loss	Validation accuracy (%)	Validation loss	Testing accuracy (%)	Testing loss
VGG16_0.0001_WD_0.5_64	99.06	0.03	97.25	0.06	98.00	0.07
ResNet_0.0001_WD_0.2_32	99.99	0	99.50	0.02	99.44	0.02
InceptionV3_0.0001_WD_0.5_32	99.34	0.02	99.50	0.02	98.96	0.04
DenseNet_0.0001_WD_0.2_64	99.93	0	99.50	0.01	99.36	0.02

DOI: 10.7717/peerj-cs.1082/table-10

From the table, it is noticed that ResNet50 and DenseNet121 have very close results and outperform VGG16 and InceptionV3. The two networks have impressive results, and both can be used in other applications. But in our work, only one is chosen to be the best-fitted network. If confusion matrix data is considered, ResNet50 misclassifies seven images whereas DenseNet121 misclassifies eight images. Thus, ResNet50 is nominated to be the most suitable network after conducting the discussed experiments and comparisons of the figures that depict the (a) learning curves and (b) confusion matrix for each one of the best-performing models. For example, Fig. 4 represents VGG16 results, whereas Figs. 5–7 represent the results for ResNet50, InceptionV3, and DenseNet121 respectively.

In our case, the false positives are represented by the top right corner of the confusion matrix figure where they refer to patients that do not have COVID but are positively classified. This could lead to a problem of shortage of medical supplies if they are hospitalized. As for the false negatives that are represented in the bottom left corner, they refer to patients with COVID that are classified as non-COVID. This could lead to a huge problem if they are not quarantined and are allowed to walk freely with crowds. Thus, the best-performing model is the one with the minimum number of FP and FN. In order to verify the validity of our findings, the results of our work is compared with several previous works. As shown in Table 11, the results of our models outperform all previously developed models included in the literature of this article.

Table 11:

Comparison with previous work.

Paper	Dataset	Models	Best result (Accuracy)
Chakraborty, Murali & Mitra (2022)	COVID-19 and Pneumonia x-rays	ResNet18	DenseNet (96.43%)
		AlexNet
		DenseNet
		VGG16
Yadlapalli et al. (2022)	COVID19-CT	VGG16	VGG16 (89%)
		ResNet50
		InceptionV3
Padma & Kumari (2020)	COVID chest Xray	2D CNN	Training ACC: 99%
			Validation ACC: 98.3%
Khan et al. (2022)	BIMCV- COVID19+	EfficientNetB1	EfficientNetB1 (96.3%)
		NasNetMobile
		MobileNetV2
Aggarwal et al. (2022)	Two datasets	MobileNetV2
		Xception
		ResNet50V2	Dataset 1:
		DenseNet121	DenseNet121 (97%)
		inceptionResNetV2	Dataset 2:
		VGG19	MobileNetV2 (85%)
		NASNetMobile
		Inceptionv3
Nazish et al. (2021)	COVID-19 patients x-rays	Logistic Regression	SVM (96%)
		SVM
Present study	COVID-19 radiography database	VGG16	ResNet50
		ResNet50	Training ACC: 99.99%
		InceptionV3	Validation ACC: 99.50%
		DenseNet121	Testing ACC: 99.44%

DOI: 10.7717/peerj-cs.1082/table-11

Discussions and limitations

Further explanations and discussions are presented. Prior research has shown that DenseNet121 is a good choice for x-ray images as noticed in our work. Furthermore, they stated that InceptionV3 is a good choice but neglected the performance of ResNet50 even though it is used in their work. This might be due to not experimenting with enough hyperparameters to reveal the quality of its performance which has been shown in our work. VGG16 got the lowest results which are normal because of its shallow architecture when compared to the other architectures.

On the other hand, InceptionV3 has 48 layers close to the number of layers of ResNet50 which is 50. Thus, our experiments have shown that the residual block technique is more useful than the inception module technique which is introduced by InceptionV3.

Finally, DenseNet121 which has 121 layers has achieved close to ResNet50 which has far fewer layers. This could be due to the overparameterization problem which means that increasing the number of layers does not necessarily improve the results. Furthermore, training 50 layers take far less time compared with training all 121 layers of the DenseNet.

For the limitations, Google Collaboratory was used in this research which offers free usage of GPU, which was needed in this research to enhance the speed of execution, but with limited time per day and a limited number of sessions in the free model. Furthermore, some configurations could not be executed due to “ResourceExhaustedError” in Fig. 8. The runtime was long even with the GPU and because our research included running a lot of experiments it as difficult to do. Even though data augmentation has shown improvements with image classification but in our case, it adversely affected the results. This could be attributed to the fact that general image augmentation techniques may not be suitable for medical images. This is an interesting research question that is worth investigating in the future.

Conclusion and future work

The purpose of this research was to train different networks on the lung X-rays dataset to classify whether the patient has COVID-19 or not, and conduct comparisons between them to help the medical field during the faced pandemic. Future work includes exploring segmentation (Zaitoun & Aqel, 2015) techniques to train the model only on the interesting and important parts of the images, which could improve the results. Furthermore, using hybrid approaches through implementing recurrent neural networks (RNN) (Sherstinsky, 2020) and transformers along with attention techniques (Vaswani et al., 2017) should be taken into consideration as it is considered a rising topic these days. Different approaches used by the researchers are noticed, but mainly most of them based their work on transfer learning. This is understandable as building a network from scratch requires a lot of data and is time-consuming. Nevertheless, with enough data, building a model from scratch is applicable and might lead to better results as it is built and trained for a specific task. Future work may include building a network from scratch and conducting a comparison between the results of transfer learning networks and the network that is built from scratch. Finally, further experiments could also be done to find the best parameters for data augmentation and assess its role in medical images.

Supplemental Information