The current era of pandemic and infectious diseases demands contemporary technologies across many industries. Modern inventions and technology have advanced significantly, primarily in disease detection, control, and influence over intelligent healthcare instruments and facilities. Artificial intelligence (AI), a rapidly developing technology, was critical in the detection of COVID-19 based on the X-ray modality. Most existing studies focus on binary classification using DL models and COVID-19 X-rays and there is a limited assessment of the strengths and weaknesses of deep learning (DL) models in multiclassification of COVID-19 using X-ray images. Therefore, this study focuses on the empirical assessment of discriminative DL models for multiclassification of COVID-19 X-rays. We suggest four phases of approaches, namely, data acquisition, preprocessing, modeling, and training for four classes of diseases (normal, pneumonia, COVID-19, tuberculosis) as well as evaluation phases in this investigation. Convolution neural network, (CNN) recurrent neural network (RNN), and multilayer perceptron (MLP) were discriminative DL models implemented. The CNN model demonstrates an efective and valuable approach for the multiclassification of diseases as classification accuracies of 0.9066, 0.6278, and 0.7652 were obtained for CNN, RNN, and MLP respectively. Discriminative DL models demonstrate the feasibility of multiclassifying COVID-19 X-rays. Implementing this approach will alleviate the burden on radiologists and other medical professionals while enhancing the precision and efectiveness of COVID-19 diagnosis.
Recently, a new virus known as COVID-19 began to infect the lungs as well as the upper respiratory
tract. On the scale of a worldwide epidemic, the incidence and mortality rates have been increasing
daily [
Technological development is changing with human evolution. The current era is one in which
demand for contemporary technologies is high across many industries. Modern inventions and
technology have advanced significantly, primarily in disease detection, control, and influence over intelligent
healthcare instruments and facilities. Artificial intelligence (AI), a rapidly developing technology, was
critical in the diagnosis of COVID-19. Previously, AI was employed in the analysis of medical images,
resulting in improved accuracy and directly or indirectly contributing to the reduction of time and labor
required for COVID-19 detection, which are significant factors. AI approaches such as deep learning
(DL) and machine learning (ML) have been increasingly important in recent years for healthcare
applications [
The primary goal of [
Each of these categories has a diferent structure and learning pattern. Researchers and experts select and use these methods even in COVID-19 classification and diagnosis without a thorough investigation of the strengths and weaknesses of these methods. This may be informed by the hyperparameter of the deep learning models or data dimensionality. In addition, most existing studies focus on binary classification using DL models and COVID-19 X-ray image datasets as radiologists’ complaints of complex traits pose dificulties in interpreting COVID-19 X-ray images. Unlike many other DL-based methods that have been proposed in the literature, the DL-based approaches in this work are proposed for multiclassification of COVID-19 based on chest X-ray images that go beyond binary classification and, in a sense, this study is very promising to address other related infectious diseases and pandemic diagnosis thereby control the spread of pandemic or infectious diseases. This study also expands the evaluation methods beyond accuracy as it focuses on the empirical assessment of discriminative DL models for multiclassification of COVID-19 X-rays. Thus, the objective of this study is to: • carries out a multiclassification of COVID-19 chest X-ray using discriminative DL models. • evaluate the performance discriminative DL models for multiclassification of COVID-19 chest
X-ray using accuracy (Acc), sensitivity (Sen), specificity (Spe), F1 measure, and confusion matrix.
This is how the rest of the paper is organized. In addition to providing a brief overview of the approach, Section 2 provides the background on the discriminative DL models for image processing, while Section 3 includes a review of the relevant works. Section 4 describes in full the methodology employed in the study. The outcomes and analysis are covered in Section 5, and the study’s conclusions and future suggestions are presented in Section 6.
This section discusses a summary of the recent literature that has chosen the discriminative DL-based
methods to carry out to carry out this work. The primary goal here is to produce a study that reviews
current advancements in DL-based COVID-19 diagnosing systems. A class of deep learning models
that is employed to deliver a discriminative function in classification tasks is called discriminative
DL. Discriminative DL models are commonly developed to enhance the ability to classify patterns by
characterizing the distributions of classes based on accessible data [
1. Convolutional Neural Network (CNN): CNN is the predominant model utilized in DL perhaps
because of the ability to incorporate many layers in the learning process. The approach has
gained widespread usage, particularly in the field of image processing, in recent years. CNN is
composed of several layers, including a convolutional layer, an activation function, pooling, and a
fully connected layer [
⎛
⎞
= ⎝ ∑︁ − 1 * + ⎠
∈
(1)
2. Multi-layer perceptron (MLP): Advancements in DL and transformer particularly, MLP models
have introduced novel network architectures for computer vision challenges. While these models
have demonstrated efectiveness in various vision tasks, such as image identification, there are
still dificulties in applying them to low-level vision [
This section provides a summary of related works that depicts the recent studies of DL models for COVID-19 diagnosis. It is presented in Table 1 that identified the models used, the contributions as well as the limitations of recent studies in this area.
COVID-19 datasets may exhibit an unequal distribution of samples across diferent classes, leading to a potential bias in the model towards the class with the highest number of samples. There is a scarcity of high-quality labeled datasets, particularly during the initial phases of a pandemic. Scarcity The model possesses unique characteris- The model was trained using a limited tics due to its pre-existing training for the dataset. The dataset’s class imbalance reidentification of diferent lung ailments. To quires attention. It focuses on binary clasaddress the issue of class imbalance, the sification. weighted loss function was employed.
The suggested CNN model, developed from scratch, achieved higher accuracy when trained on a specific medical dataset compared to pre-trained models based on a more general ImageNet dataset. The model was trained using limited resources and within a short timeframe.
The COVID-Net architecture was made publicly available for open access. Utilized lightweight design patterns. Implemented selective long-range connectivity in strategic areas to enhance representational capacity and streamline training, while ensuring optimal computational complexity and memory eficiency.
Regularization techniques were employed to address the issue of data imbalance. Preprocessing the data enhanced the performance of the model.
The utilization of pre-trained networks has simplified the process of classification tasks. The calculation of pandemic uncertainty was performed to ensure the accurate identification of classes.
Utilizing preprocessing procedures (histogram equalization algorithm, and a bilateral low-pass filter) efectively to improve the performance of CNNs in the detection of COVID-19 from chest X-ray.
The model underwent training using a restricted dataset. The study was limited to two classes.
Enhancing sensitivity is crucial to minimize the number of COVID-19 cases that go undetected. It focuses on binary classification.
The model sufers the issue of class imbalance. It focuses on binary classification.
Exclusively concentrated on the posterioranterior (PA) perspective of the X-rays, hence unable of distinguishing alternative viewpoints of the X-ray images. The X-ray images containing multiple disease symptoms were not eficiently classified. It focuses on binary classification.
Possible constraints on the capacity to apply findings to a wider context, resulting from particular methods used to prepare the data and the unique attributes of the dataset.
The utilization of batch normalization ap- Enhancing sensitivity is necessary to miniproach enhances the rate at which conver- mize the number of undetected COVID-19 gence occurs during the training process. instances. The study was limited to binary The vanishing gradient problem was re- classification. solved through the utilization of a residual network. can result in overfitting, a situation where the model exhibits good performance on training data but performs badly on unseen data. Most importantly, limited studies focus on the performance of the discriminative DL models on multi-classification of X-ray images using the COVID-19 dataset. Rather, the focus has been on binary classifications of the COVID-19 X-ray. In addition, there are limited studies on multiclassification of pandemics and infectious as most of the existing studies focus on binary classification.
The current study expands the investigation on COVID-19 detection beyond normal (healthy) vs
infected (unhealthy) using chest X-ray images, the author explores a suitable X-ray dataset that has
four diferent cases (tuberculosis, pneumonia, COVID-19 and normal) to investigate the robustness of
discriminative DL model in the multiclassification of COVID-19 X-ray dataset. The study methodology
phases align with some elements of CRoss Industry Standard Process for Data Mining (CRISP-DM)
that proposed a systematic approach that categorizes operations into four levels of abstraction: phases,
generic tasks, particular tasks, and processes integrated into the life cycle of data mining projects [
The data for this study is a secondary X-ray image dataset from a popular repository named, Kaggle, https://www.kaggle.com/datasets/jtiptj/chest-xray-pneumoniacovid19tuberculosis. It was titled Chest X-Ray (Normal/Pneumonia/Covid-19/Tuberculosis). The content of the dataset is structured into three (3) folders (train, test, val) and contains subfolders for each image category or class (Normal, Pneumonia, COVID-19, and Tuberculosis), this makes it useful for multi-classification studies. The creation of the dataset was inspired by AI applications, specifically to detect and classify these diseases from X-ray images. A total of 7135 x-ray images are present. Figure 3 presents eight (8) samples of the acquired dataset. Images 1, 4, 5, and 8 show pneumonia infection, image 2 shows tuberculosis infection, images 3 and 7 are normal chest X-ray images while image 6 shows the COVID-19 infection dataset. The websites from where the X-ray images were taken asserted that they had addressed the ethical issues related to obtaining and utilizing the images.
The images acquired exhibit inconsistencies due to the dataset being obtained from multiple origins employing diverse equipment and attributes. Therefore, it is necessary to do pre-processing on the data before feeding it to the model for further processing. To make the input chest X-ray images compatible with the three DL-based models considered for empirical assessment in this study, images were first scaled to 224 × 224 pixels. However, to enhance the algorithm’s ability to generalize, this research refrains from using substantial pre-processing processes. Therefore, three pre-processing steps were used for the training process:
Random flip is an example of augmentation, it was used to generate variants of x-ray image dataset for this study. This is medical data so not much augmentation should be done to avoid misrepresenting the data and having false results. The random flip seems to be something easy that can be done without much efect.
A consistent 224 x 224-pixel resizing is needed and applied to every image.
Min-max normalization was employed to ensure consistent intensity across all photos. The intensity value of all images, in the range ([0.485, 0.456, 0.406]), was normalized to the intensity range of 224 × 224 following equation 2.
=
− ( ) ( ) − ( ) (2)
Various kernel functions were utilized in CNN, RNN, and MLP models to classify the acquired dataset for COVID-19 classification, employing a distinct approach. The models for COVID-19 detection were subsequently trained (fine-tuned) using chest X-ray images. The model was trained for 28 epochs. The initial optimizer used was the Adam Optimiser and the CrossEntropy loss was the loss function that was used for the training process. The architecture including a total of 23 layers was built for the investigation. Python programming was employed in the software development process, where the names of functions and parameters for the layers were directly specified in the respective sections for layer names and parameters. The presence of four dimensions in the first layer, known as the picture input layer, is a result of conducting trials involving various input photos of varying sizes as part of the study. Figures 4 - 6 show the parameters setting for the CNN, RNN, and MLP models’ architecture respectively.
The evaluation criteria proposed for the empirical assessment of the CNN, RNN, and MLP in the
multiclassification of COVID-19 X-rays are Acc, Sen, Spe, F1 measure, and confusion matrix. This study
utilizes various parameters, including false positive (FP), true positive (TP), true negative (TN), and
false negative (FN), as well as dependent variables such as Acc, Sen, Spe, F-1 measure and confusion
matrix. These variables were obtained through mathematical analysis of the aforementioned parameters
and were used to evaluate the results. In this case, TP represents the frequency with which the real
patient data is correctly identified as patients by classification. False positive (FP), on the other hand,
refers to the instances where non-patient data is incorrectly classified as patient data. TN represents
the frequency at which data that is not related to patients is incorrectly identified as not being related
to patients due to the classification process. FN, however, involves categorizing the patient data as
non-patient data in a similar manner. The mathematical definitions of Acc, Pre, Sen, Spe, F-1 measure
and confusion matrix, values obtained using these parameters are given in equations 3 – 7. Spe, F-1
score, and Acc were computed using a threshold (cut-of) value of 0.5, this is according to [
+ + + + =
+ (4) = =
+
+ * 1 − = 2 * + (5) (6) (7)
This section presents the results attained by three types of Neural Network architectures (discriminative DL models) on the joint diagnosis of COVID-19, tuberculosis, and pneumonia. The task is conceptualized as a multiclass classification problem on chest X-ray images. We also discussed and analyzed the outcome of this investigation in this section. Three discriminative DL models are investigated in the COVID-19 X-ray image dataset. Section 5.1 presents the outcome of the investigation while Section 5.2 presents the analysis of the outcome.
The investigation results are based on the multi-classification discussed in Section 5.1. Table 2 shows the performance of the three discriminative DL over the COVID-19 X-ray dataset earlier discussed in
Pre 1.0000 0.8978 0.9214 0.7018
Section 5.2 in the empirical assessment study. Figure 7 depicted the Loss vs Epoch curve for training and validation.
These findings emphasize the capabilities of discriminative DL models in enhancing X-ray diagnosis and clinical environments. Upon analyzing the outcomes of the empirical assessment of the discriminative DL model for the multiclassification of COVID-19 X-ray presented in Table 2, it becomes evident that each of these discriminative DL models has areas of strength as well as weaknesses, depicting the average performance of the models, and it shows that CNN significantly outperformed the two other discriminative DL models. Figure 7 displays the curve depicting the relationship between loss and epoch for the training and validation of a COVID-19 X-ray CNN. The training loss exhibits a consistent decline with each subsequent epoch, suggesting that the model is progressively acquiring knowledge and efectively adapting to the training data.
There is a significant and rapid decrease in the early stages, followed by a steady slowdown, indicating that the model immediately grasps the fundamental patterns and then refines them. Similarly, the validation loss first lowers but then levels of after a few epochs. The validation loss remains very steady with slight fluctuations after the initial decline, suggesting that the model is efectively adapting to the unseen validation data. The persistent decrease in training loss demonstrates the successful acquisition of knowledge from the training data. The consistent validation loss indicates that the model is not sufering from overfitting and is efectively adapting to new data.
The curve illustrating the correlation between loss and epoch for the training and validation of the COVID-19 X-ray RNN is shown in Figure 6. The training loss consistently declines from one epoch to the next, indicating that the model is gaining knowledge and improving its performance on the training data. The fall in the second graph is less pronounced than in the first graph, indicating a positive trend in learning progress. The validation loss has a reasonably low value in comparison to the training loss, indicating that the model is not sufering from overfitting and is efectively generalizing to the validation data. The validation loss exhibits slight changes but maintains a generally constant trend. The training loss exhibits a decreasing trend, but towards the end, it shows a little increase, maybe indicating the onset of overfitting. It is something that should be observed in future periods. The low and stable validation loss signifies a commendable level of generalization performance.
Figure 8 displays the curve depicting the relationship between loss and epoch for the training and validation of the COVID-19 X-ray MLP. The first training loss exhibits a significant value, but swiftly diminishes within the initial epochs, suggesting that the model promptly grasps the essential patterns inside the training data. Following the first decline, the training loss consistently decreases, indicating that the model is gradually enhancing its alignment with the training data. The validation loss has an initial low value and maintains a consistently low level throughout the epochs, with minimal variations. The model’s validation loss demonstrates stability and a low value, indicating that it is efectively generalizing to the validation data and avoiding overfitting. The sharp and rapid decrease in training loss indicates that the model is quickly acquiring the fundamental characteristics of the data. The consistent and little validation loss seen during the training phase suggests successful generalization and the absence of overfitting.
Figures 10 - 12 are the confusion matrixes for CNN, RNN, and MLP respectively in this multiclassification study. The models demonstrate rapid acquisition of features and accurate X-ray image classification after only a few epochs of training. In addition, this slightly parallels the CNN model. This means CNN and the MLP are the best at the classification problem. But in all, CNN proved to be an outstanding model among the three discriminative DL explored in this study.
The goal of the research was to empirically evaluate the strengths and weaknesses of discriminative DL models in a multiclassification task using a dataset of COVID-19 X-rays, with the aim of achieving accurate results for disease diagnosis. This study expands the investigation of discriminative DL models over the COVID-19 X-ray dataset beyond binary classification tasks to multiclassification tasks due to the evolution of related infectious diseases. The detection models built using CNN, RNN, and MLP models were able to establish that the discriminative DL models can perform multiclassification tasks using COVID-19 X-ray datasets (tuberculosis, pneumonia, COVID-19, and normal) which are complex traits and pose dificulties for radiologists to interpret. Diferent preprocessing techniques such as data normalization, resizing, and random flip flop were used. The high accuracies achieved indicate that the discriminative DL models can identify unique features in the COVID-19 X-rays, enabling the deep networks to accurately diferentiate between the images. These trained models may efectively reduce the workload of radiologists and other medical practitioners and increase the accuracy and eficiency of COVID-19 diagnosis. The CNN model demonstrates an efective and valuable approach for the multiclassification of diseases. This approach for identifying individuals with COVID-19 using chest X-ray pictures can be employed for preliminary screening purposes. To enhance the performance of the developed model, it is advisable to train the models on diferent datasets other than chest X-ray as this may be the subject of comparison to improve these models in the future, in a sense, these models in this research can be trained on an alternative dataset (CT scan, audio etc) to acquire a more profound understanding of the performance. Researchers can also explore the performance of several CNN architectures, such as AlexNet, ResNet50, and InceptionV3, and utilize metaheuristic optimization algorithms may be explored to set hyperparameter turning of DL network and improve the data exploration for such experiments in the future. These models and architecture can also be trained on a multimodal COVID-19 dataset to get a reliable diagnosis of the infection. A real-time scalable with an IoT framework COVID-19 diagnostic program can be created based on existing research findings to facilitate its implementation in medical practice.
Acknowledgments: Authors acknowledge the Centre of Excellence, University of Zululand for the support received for this work.