Detection and Analysis of Covid-19 in Chest X-Ray Images: A Review

Abstract

For suspected instances of coronavirus illness, chest X-ray (CXR) imaging is a routine and critical examination tool (COVID-19). Because of its availability, cheap cost, and quick findings, CXR imaging is preferred in severely damaged or resource-limited locations. But, considering COVID-19\'s fast spread, such testing may restrict the effectiveness of pandemic control and prevention. Ai and automation such as machine learning, which have reached province achievement in the interpretation of visual input and a wide variety of clinical pictures, are pledging choices for automatic treatment in response to the problem. This research examines and evaluates preliminary and scientific articles for the identification of COVID-19 using CXR pictures using dcnns and other machine learning designs between March and May 2020. Given the promising results, public, complete, and varied datasets are urgently needed. For more robust, honest, and accurate recommendations, additional research in terms of explainable and justified judgments is also necessary.

Introduction

I.INTRODUCTION

Early detection of hemorrhagic fever disease (COVID-19) is critical for limiting virus transmission and providing treatment to avoid consequences. The daily increases in COVID-19 cases throughout the world, as well as the limits of current clinical diagnosis, make recognising and controlling the pandemic difficult. Researchers from all around the globe are working together to create more effective diagnostic techniques and to speed up the development of a vaccination and therapies. Three diagnostic methods are routinely utilised as of the writing of this paper: blood testing, virus tests, and radiology [1]. Antibodies to the middle east respiratory syndrome coronavirus 2 (SARS-CoV-2) are found in the blood-by-blood testing. However, this test's accuracy in identifying COVID-19 is as low as 2% or 3% [2]. Samples from the respiratory tract are used in viral testing to identify SARS-CoV-2 antigens. The rapid diagnostic test (RDT) is a type of antibody detection test that can provide findings in as little as 30 minutes. However, RDT test kits are few, and their efficiency is dependent on the quality of the sample and the beginning of disease. However, because the test does not discriminate COVID-19 from other viral infections, it might provide false positive findings; as a result, it is not advised for diagnosing COVID-19 [3]. The polymerase chain reaction is another regularly used viral diagnostic (RT-PCR). The gold-standard tool for first-line screening is RT-PCR. Large-scale investigations, on the other hand, have indicated that the test result sensitivity varies from 50 to 62 percent [4]. This means that a negative RT-PCR result can be achieved at the outset. Multiple RT-PCR tests are done during a 14-day monitoring period to guarantee the accuracy of the test result for diagnosis. In plenty of other words, a true negative RT-PCR result for a suspected case of COVID-19 is only regarded when no significant RT-PCR findings are found after repeated tests have been performed throughout the 14-day evaluation period [5]. Due to a lack of RT-PCR test kits in certain countries, this can be aggravating for patients and costly for healthcare providers [6].

COVID-19 attacks the respiratory system, therefore chest radiography images are crucial for detection and better treatment. In Italy and other countries, chest X-rays (CXR) have been utilised as a first-line diagnostic technique [7]. The state of the lungs, as well as the various stages of disease or recovery, may be accurately diagnosed utilising radiological scans [8]. COVID-19 patients' radiological scans revealed a variety of anomalies, according to radiologists. In CXR images, COVID-19 characteristics such as bilateral GGO and reciprocal and multifocal GGO with condensation .

CXR is a frequently used technology in most hospital studies; it takes less time to prepare patients and provides quick results. As a result, CXR may be used to triage patients, prioritise patient therapies, and allocate healthcare supplies.

Deep learning (DL) approaches have been utilised to dramatically increase image processing performance in the medical imaging area [9], [10]. Microscopy photos [11], brain tumour categorization [12], MRI images [13], and retinal photographs [14] are just a few examples of where DL has been used effectively.

Convolutional neural networks (CNNs) are widely employed in medical imaging [15], [16], and come in a variety of architectural and application forms. As a result, since the beginning of the epidemic, DL techniques for identifying COVID-19 from radiological pictures have been extensively investigated. We compare the existing DL latest technology, identify the issues, and find the necessary future exploration in this paper, which reviews the latest research charitable donations of the proposal of DL for the diagnosis of COVID-19 from CXR images by considering the actual DL latest technology, outlining the challenges, and recognising the continuous new exploration.

This research examines and critically evaluates the preprint and published papers on COVID-19 diagnosis using CXR pictures that were made public between March and May 2020 to get a better understanding of how CNNs and other DL architectures may aid in the diagnosis of COVID-19 via CXR images. The publications were located in PubMed, ScienceDirect, Springer, IEEE, ACM, Scopus, ArXiv, and MedRxiv, among other research databases. "Transfer learning," "convolutional," "deep learning," "radiograph," "chest x-ray," "CXR," "COVID," and "Coronavirus," among the terms used in the search;" Since the commencement of the study, this list has been updated on a regular basis. We looked through the abstracts and ruled out research that utilised typical machine learning techniques and explored deep learning for computed tomography images. Only the most current pieces from several resources were evaluated where they overlapped.

II. RELATED WORK

We looked at 34 papers that looked at how DL models were used to evaluate CXR pictures with SARS-CoV-2 viral infections. The majority of the research (71%) used publicly accessible CNN architectures trained on the ImageNet dataset to apply transfer learning. The parameters and fine - tuning settings for these designs are publicly accessible [17]. However, only 29% of the research used off-the-shelf technologies and instead used new structures. We offer a basic analysis of the main methodologies and datasets employed in the research papers covered in this assessment in the sub groups that follow.

A. Classification Task Formulation

COVID-19 detection findings are obtained by categorising CXR pictures into two to four classes, i.e. binary or multi-class identification. "Healthy," "no finding," "bacterial pneumonia," 'viral pneumonia," or "COVID-19," each class contains one or more labels: "strong," "no trying to find," "microorganisms pneumonia," "viral bronchitis," or "COVID-19." The COVID-19 designation and one of the following labels: "healthy," "no finding," "microbial sinusitis," or "viral bronchitis," are the findings of binary classification, which has two classes. "COVID-19," "sound or no finding," and "pneumonia" are among the three classes of outcomes. "COVID-19," "healthy or no finding," "bacteremia," and "viral tuberculosis," are the four classes" outcomes. The majority of the studies examined employed two or three courses. The number of evaluated studies is categorised by the number of classifying classes utilised in the assessment job in Figure 1

B. Datasets

14 distinct datasets were utilised in the peer-reviewed studies. COVID-19 Image Data Collection [18] is the most often mentioned dataset, according to our poll. It comprises photos taken from a variety of online publications and websites in an attempt to give COVID-19 photographs to AI researchers for the development of deep learning-based models. This dataset includes a collection of variables for each photograph, including sex, age, date, survival, and clinical remarks.

C. Transfer Learning

In medical imaging applications, transfer learning is commonly used [15], [17]. When there aren't enough training instances to train a model from start, ensemble learning comes in handy. Tajbakhsh et al. [16] shown that a fine-tuned well before CNN can outperform or equal a CNN trained from zero. As a result of the restricted training data, transfer learning has been investigated intensively for COVID-19 identification from CXR pictures.

The examined works that used transfer learning may be divided into three categories in this survey. In the first group, the values of a human brain that would be trained on the target CXR data were initialised using an or before CNN on a large-scale natural picture dataset. Models trained on ImageNet, for example, were utilised in [16], [17], and [18]. The second category of research includes those in which the weights of some of the early layers of a pre-trained model on a large-scale natural picture collection were frozen while the last layers were finetuned [19].

This is because early-layer characteristics (such as edges) are more generic, but later-layer features are more unique to a certain job or dataset [17]. a few examples

D. CNN Architectures

CNN architectures [9] have recently achieved human expert-level performance in a variety of demanding visual tasks, including as medical picture evaluation and pathology identification. Since the first successful CNN in 1998, other CNN designs have been described in the literature. It was widely used for handwritten digit recognition [22]. It was known as LeNet and was created by Yann LeCun. LeNet has a shallow design compared to existing models, with three volumetric, two average pooling, and upsampling. The CNN architectures employed in the evaluated research, as well as their usage and outcomes for COVID-19 identification from CXR pictures, are briefly described in the subsections below.

1. AlexNet: AlexNet [13] is similar to LeNet, but it is more complex, with three convolutional layers layered on top of each other. AlexNet took first place in the 2012 ILSVRC competition, with a top-five error rate of 17%. The authors employed a dropout regularisation approach and data augmentation in AlexNet to tackle the overfitting problem. Razzak et al. [14] utilised AlexNet to classify COVID-19 instances into binary and multiclass categories. COVID-19/healthy binary class accuracy was 97.04 percent, while COVID-19/healthy/bacterial pneumonia/viral pneumonia multiclassification success was 63.27 percent. Kumar and Kumari achieved a 93.0 recognition rate using AlexNet as a learning algorithm to feed a gradient boosting machine (SVM) classifier. AlexNet was also utilised by Abbas et al. [16] to extract features from three classes: normal, COVID-19, and SARs. However, they suggested that a class breakup layer be created to divide each class into several sub-classes in their paper. The final forecasts were created by reassembling these subsets. They employed AlexNet to discover features for the suggested decomposed layer, and their proposed model DeTraC had a 95.12 percent accuracy.
2. GoogleNet: The 2014 ILSVRC competition was won by Google Net [17], which had a top-five mistake rate of 6.67 percent. In addition, towards the pooling and convolutional layers, GoogleNet includes an iterative algorithm, making it substantially more complex than earlier CNNs (IM). This unit can learn bridge correlates (depth-wise) as well as spatial correlation by acting as a tiny network. There are six convolutional neural networks in total: four 1x1 layers, one max pooling layer, and one synthesis layer [17]. The IM acts as a bottleneck layer and provides a number of benefits. For starters, it allows for the training of far deeper models while lowering the number of training examples by approximately tenfold. Second, an IM's output is set to be less than its intake in terms of number of items. As a result, the IM diminishes complexity. Third, together with the spatial and depth dimensions, an IM may capture complicated patterns at numerous scales. Other Google Net variations have been developed, each with somewhat different inception ingredients, and they have outperformed the original. Inception-V2, Inception-V3 [18], Inception-ResNet, and Inception-v4 [9] are other examples.
3. VGGNet: The Visual Geometry Group (VGG) at Oxford University submitted VGGNet [4], which came in second place in the 2014 ILSVRC project with a top-five error rate of 7.3 percent. VGGNet has the benefit of architectural simplicity [4], with a total of 16 or 19 convolutional layers. It did, nonetheless, employ three times as many parameters as AlexNet [18].
4. ResNet: The CNN architecture was enhanced by ResNet [13], which included a deep residual feature. A normal layer with a skip link makes up the residual unit (RU). The skip service allows a layer's input signal to transit the net by attaching it to that layer's output. As a result of the RUs, an extraordinarily deep model with 152 layers was trained, which won the 2015 ILSVRC competition and had a top-five error rate.
5. Xception: Francois Chollet presented Xception [15], which outperformed Inception-V3 on a large image classification assignment with 350 million pictures and 17,000 classes. Unlike GoogleNet's IM, the basic principle of Xception is to learn both cross-channel and geographic distribution individually. The IM is thus supplanted in Xception by an outstanding termed depth-wise separate compression. When opposed to Inception-V3 (78.2 percent accuracy) on the ImageNet dataset, the differentiated convolution operation had fewer features, lower processing cost, and overall greater accuracy of 79 percent [15].
6. SENet: Hu et al. [17] proposed the Squeeze and Excitation Network (SENet), which won the 2017 ILSVRC Challenge with a top-five error rate of 2.251 percent. SENet improves the performance of the GoogleNet IM and ResNet designs by extending them. The SE block was introduced by SENet as a new component. This element has been enabled to each ResNet IM or RU. SE-Inception and SE-ResNet are the expanded versions, respectively. Three basic layers help compensate the SE block: a layer for worldwide mean pool across feature maps, a squeeze dense layer, and a dense layer.

Figure 1 In the reviewed study, Deep learning architectures were employed.

III. METHODOLOGY ANALYSIS

Various DL architectures (CNNs in especially) have indeed been presented for the identification of COVID-19 from CXR pictures in a reasonably short period, as stated in earlier sections. This section contains in-depth information regarding the studies that have been examined.

A. Datasets

Except for the research of Iqbal Khan et al. [9] and Gomes et al. [14], all of the investigations assessed employed accessible to the public information.

To expand the training set, most of the research pooled datasets. However, one of the primary obstacles is an imbalances problem caused by the restricted number of COVID-19 samples available.  Most of the research used a variety of architectures to compare classification results or develop an ensemble model to improve efficiency.

B. Deep Learning Models Constructing

Most researchers favoured transfer learning, according to the review, and there is still strong interest in this method. Transfer learning, in example, allows for quick model construction while surpassing other methods. The most often used pre-trained models are ResNet [13], DenseNet [18], Inception [18], and VGG [20]. A transfer learning model that has been adequately educated will typically outperform a model that has been taught starting zero. The greater the quality, the smaller the dataset. However, because the supply of labelled COVID-19 CXR pictures is currently restricted, or before models with thousands of parameters, such as VGG or ResNet, can easily overfit the training examples. As a result, special emphasis must be given to obtaining relevant and representative testing data as well as identifying relevant and representative metrics for review.

C. Performance Comparisons

Due to differences in the size of the training images and the lack of uniform quality metrics, it was impossible to compare the research that are included in the study, further complicating the selection of the most productive DL models for identifying COVID-19 from CXR pictures. The efficiency, sensitive, and specificity criteria were used by the majority of writers to assess the DL models. When non-standard measurements and information from diverse sources are employed, however, assessing alternative methodologies becomes more challenging. As a result, creating a public COVID-19 dataset that is complete and approachable to the AI research community is critical. Furthermore, criteria for analyzing the quality of estimation techniques must be created..

IV. DISCUSSIONS

Despite the positive outcomes of the DL architectures, there are a number of challenges that need to be addressed in order to improve the diagnostic process' accuracy, transparency, and trustworthiness. The current research issues connected with the identification of COVID-19 from CXR images are highlighted in this section. To increase the number of photos retrieved from the restricted number of COVID-19 instances, image enhancement was used. Multiple dataset are made up of photos with unrelated visual characteristics and deceptive artefacts that aren't generally tackled by the research in this study.]

A. Class Imbalance Problem

The COVID-19 datasets have a problem with class imbalance. The unequal distribution of classes raises questions about the machine learning algorithm's resilience. Kumar et al. [7] advocated the use of SMOTE to alleviate this issue in one of their investigations. Ucar and Korkmaz [18] and Rajaraman and Antani [14] both advocated for the establishment of a strategy.

B. Explaining Deep Model Predictions

The majority of the research in this study employ DL architectures as black-box classifiers, with no explanation of the model decisions. Explainable AI is a new AI area that refers to strategies and methods for deciphering the decision-making processes made by machine learning models. Wang and Wong [1] employed the GSInquire technique [92] to identify the places where the DL classifier is used to drive recommendations.

C. Managing Classification Uncertainty

Uncertainty in DL refers to the degree of confidence in the classifier's result [3]. Even though the softmax output might be confused with model certainty [1] [10], having a high softmax output does not imply great certainty. Uncertain instances can be dealt with caution using a DL model that acknowledges uncertainty. When a model produces a result with a high level of uncertainty, it is suggested that human interaction be used to investigate the result further [8].

D. Covid-19 Severity Assessment

Increased infection assessments and prognosis analysis are two more issues that have yet to be addressed in the COVID-19 CXR imaging literature. CXR imaging analysis might also aid in the identification of high-risk individuals and regions that require immediate attention and assistance. These concerns and problems necessitate greater medical engagement at all phases of the creation, assessment, and validation of DL models. Due to the expanding number of patients who require prompt and correct critical care and resources during the COVID-19 pandemic, triage is a crucial step. COVID-19 patients are easily triaged thanks to DL studies that try to anticipate, follow, and measure their progress and severity. Tracking the development of COVID-19 patients was considered by Duchesne et al. [5] and Islam and Fleischer [7].

Conclusion

This research provides a thorough examination of a wide range of topics. The majority of research utilising the same dataset obtained by Cohen et al. [18] employed CNN-based transfer learning. Spite of the good results gained, there is still a lot of potential for development. First and foremost, public, comprehensive, and diversified databases must be created. Authorities should evaluate the datasets and label them with the associated lung disease lesions. The prediction accuracy and model transparency would both improve if the detection of indicators was included with the categorization output. Second, because medical research into COVID-19\'s major characteristics is still ongoing, it is critical to use additional features extracted based on medical personnel\'s suggestions. Given the limited quantity of accessible CXR COVID-19 datasets, including domain knowledge would aid in the development of models that replicate human expert diagnostic patterns and concentrate on the indications or areas that they pay special attention to. However, adequate domain knowledge must be established first. To achieve the desired performance, the trade-off between the automatically learnt deep features and the extracted domain knowledge features should be handled. Third, in order to build a benchmark for use in the prediction assessment of deep learning models, it is necessary to evaluate the level of discrepancy amongst radiologists. Fourth, because physicians frequently resort to prior similar instances to make trustworthy diagnoses, we believe that semi-supervised learning has a lot of promise that has yet to be realised. As part of the training set, semi-supervised algorithms use a small number of labelled samples and a large amount of unlabeled data. Semi-supervised modelling can assist uncover hidden patterns and relationships in data, as well as lower the cost of data annotation. Fifth, as can be observed, typical data augmentation methods were used in majority of the investigations in our review to deal with the lack of COVID-19 CXR pictures. The promising outcomes of generative adversarial networks (GAN) should be investigated further. Finally, the encouraging results obtained by utilising GenSynth to automatically generate a deep CNN architecture optimised for the COVID-19 classification challenge may be shown.

References

[1] World Health Organization. (2020). Guidance and Standard Operating Procedure: COVID-19 Virus Testing in NHS Laboratories. Accessed: Jun. 5, 2020. [Online]. Available: https://www.england.nhs. uk/coronavirus/publication/guidance-and-standard-operating-procedure-covid-19-virus-testing-in-nhs-laboratories/ [2] World Health Organization. (2020). WHO Warns That Few Have Devel oped Antibodies to Covid-19. Accessed: Jun. 5, 2020. [Online]. Avail- able: https://www.theguardian.com/society/2020/apr/20/studies-suggest- very-few-have-had-covid-19-without-symptoms [3] World Health Organization. (2020). Advice on the Use of Point-of-Care Immunodiagnostic Tests for COVID-19. Accessed: Jun. 5, 2020. [Online]. Available: https://www.who.int/news-room/commentaries/detail/advice- on-the-use-of-point-of-care-immunodiagnostic-tests-for-covid-19 [4] J.-L. He, L. Luo, Z.-D. Luo, J.-X. Lyu, M.-Y. Ng, X.-P. Shen, and Z. Wen, ‘‘Diagnostic performance between CT and initial real-time RT- PCR for clinically suspected 2019 coronavirus disease (COVID-19) patients outside Wuhan, China,’’ Respiratory Med., vol. 168, Jul. 2020, Art. no. 105980 [5] E. B. G. Kana, M. G. Z. Kana, A. F. D. Kana, and R. H. A. Kenfack, ‘‘A Web-based diagnostic tool for COVID-19 using machine learning on chest radiographs (CXR),’’ medRxiv, Jan. 2020, Art. no. 20063263. [6] H. Y. F. Wong, H. Y. S. Lam, A. H. T. Fong, S. T. Leung, T. W. Y. Chin, C. S. Y. Lo, M. M.-S. Lui, J. C. Y. Lee, K. W. H. Chiu, T. W. H. Chung, and E. Y. P. Lee, ‘‘Frequency and distribution of chest radiographic find- ings in patients positive for COVID-19,’’ Radiology, vol. 296, Mar. 2020, Art. no. 201160. [7] A. E. Hassanien, L. N. Mahdy, K. A. Ezzat, H. H. Elmousalami, and H. A. Ella, ‘‘Automatic X-ray COVID-19 lung image classification sys- tem based on multi-level thresholding and support vector machine,’’ medRxiv, Mar. 2020, Art. no. 20047787. [8] T. Ozturk, M. Talo, E. A. Yildirim, U. B. Baloglu, O. Yildirim, and U. R. Acharya, ‘‘Automated detection of COVID-19 cases using deep neural networks with X-ray images,’’ Comput. Biol. Med., vol. 121, Jun. 2020, Art. no. 103792. [9] G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A. W. M. van der Laak, B. van Ginneken, and C. I. Sánchez, ‘‘A survey on deep learning in medical image analysis,’’Med. Image Anal., vol. 42, pp. 60–88, Dec. 2017. [10] F. Altaf, S. M. S. Islam, N. Akhtar, and N. K. Janjua, ‘‘Going deep in medical image analysis: Concepts, methods, challenges, and future directions,’’ IEEE Access, vol. 7, pp. 99540–99572, 2019. [11] F. Xing, Y. Xie, H. Su, F. Liu, and L. Yang, ‘‘Deep learning in microscopy image analysis: A survey,’’ IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 10, pp. 4550–4568, Oct. 2018. [12] K. Muhammad, S. Khan, J. D. Ser, and V. H. C. de Albuquerque, ‘‘Deep learning for multigrade brain tumor classification in smart healthcare systems: A prospective survey,’’ IEEE Trans. Neural Netw. Learn. Syst., pp. 1–16, 2020, doi: 10.1109/TNNLS.2020.2995800. [13] J. Liu, Y. Pan, M. Li, Z. Chen, L. Tang, C. Lu, and J. Wang, ‘‘Applications of deep learning to MRI images: A survey,’’ Big Data Mining Anal., vol. 1, no. 1, pp. 1–18, Mar. 2018. [14] P. Seebock, J. I. Orlando, T. Schlegl, S. M. Waldstein, H. Bogunovic, S. Klimscha, G. Langs, and U. Schmidt-Erfurth, ‘‘Exploiting epistemic uncertainty of anatomy segmentation for anomaly detection in retinal OCT,’’ IEEE Trans. Med. Imag., vol. 39, no. 1, pp. 87–98, Jan. 2020. [15] H.-C. Shin, H. R. Roth, M. Gao, L. Lu, Z. Xu, I. Nogues, J. Yao,D. Mollura, and R. M. Summers, ‘‘Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteris- tics and transfer learning,’’ IEEE Trans. Med. Imag., vol. 35, no. 5, pp. 1285–1298, May 2016. [16] S. M. Anwar, M. Majid, A. Qayyum, M. Awais, M. Alnowami, and M. K. Khan, ‘‘Medical image analysis using convolutional neural net- works: A review,’’ J. Med. Syst., vol. 42, no. 11, pp. 1–13, Nov. 2018. [17] R. Yamashita, M. Nishio, R. K. G. Do, and K. Togashi, ‘‘Convolutional neural networks: An overview and application in radiology,’’ Insights Imag., vol. 9, no. 4, pp. 611–629, Aug. 2018. [18] J. Paul Cohen, P. Morrison, L. Dao, K. Roth, T. Q Duong, and M. Ghassemi, ‘‘COVID-19 image data collection: Prospective predic- tions are the future,’’ 2020, arXiv:2006.11988. [Online]. Available: http://arxiv.org/abs/2006.11988 [19] (2021). Radiopaedia.org. Accessed: Jan. 26, 2021. [Online]. Available: https://radiopaedia.org/ [20] M. E. H. Chowdhury, T. Rahman, A. Khandakar, R. Mazhar, M. A. Kadir, Z. B. Mahbub, K. R. Islam, M. S. Khan, A. Iqbal, N. Al-Emadi, M. B. I. Reaz, and T. I. Islam, ‘‘Can AI help in screening viral and COVID-19 pneumonia?’’ 2020, arXiv:2003.13145. [Online]. Available: http://arxiv.org/abs/2003.13145 [21] X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers,‘‘ChestX-ray8: hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax dis- eases,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017, pp. 2097–2106. [22] P. Mooney. Chest X-Ray Images (Pneumonia) | Kaggle. Kaggle.com.Accessed: 2018. [Online]. Available: https://www.kaggle.com/ paultimothymooney/chest-xray-pneumonia

Copyright

Copyright © 2022 Sheikh Raheela Gaffar, Mr. Ankur Gupta. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.