Truncated Inception Net: COVID-19 Outbreak Screening using Chest X-rays

Abstract

Since December 2019, the Coronavirus Disease (COVID-19) pandemic has caused world-wide turmoil in less than a couple of months, and the infection, caused by SARS-CoV-2, is spreading at an unprecedented rate. AI-driven tools are used to identify Coronavirus outbreaks as well as forecast their nature of spread, where imaging techniques are widely used, such as CT scans and chest X-rays (CXRs). In this paper, motivated by the fact that X-ray imaging systems are more prevalent and cheaper than CT scan systems, a deep learning-based Convolutional Neural Network (CNN) model, which we call Truncated Inception Net, is proposed to screen COVID-19 positive CXRs from other non-COVID and/or healthy cases. To validate our proposal, six different types of datasets were employed by taking the following CXRs: COVID-19 positive, Pneumonia positive, Tuberculosis positive, and healthy cases into account. The proposed model achieved an accuracy of 99.96% (AUC of 1.0) in classifying COVID- 19 positive cases from combined Pneumonia and healthy cases. Similarly, it achieved an accuracy of 99.92% (AUC of 0.99) in classifying COVID-19 positive cases from combined Pneumonia, Tuberculosis and healthy CXRs. To the best of our knowledge, as of now, the achieved results outperform the existing AI-driven tools for screening COVID-19 using CXRs.

1. Introduction

Coronavirus disease 2019 (COVID-19) is an infectious disease caused by Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) [1]. The disease was first identified in 2019 in Wuhan, China, and has since spread globally, resulting in the 201920 Coronavirus pandemic [2]. With more than 0.71 million cases of infection and 33,500 cases of death by the third month of its discovery (as on March 30, 2020), the SARS-CoV-2 continues to infect people worldwide [3]. The virus is primarily transmitted among individuals through respiratory droplets. Studies have also shown that the virus can persist on surfaces where an infected individual might have touched. As a consequence, as of now, the spread of this virus is exponential [3].

The gold standard for the diagnosis and detection of COVID-19 is the polymerase chain reaction (PCR). It can detect the SARS-CoV-2 RNA from respiratory specimens through na- sopharyngeal or oropharyngeal swabs. Despite the high sensitivity and accuracy of the PCR technique, the method is highly time-consuming and resource-intensive. Therefore, considering the unprecedented spread rate of the virus across the globe and the rapid temporal progression   of the disease throughout a subject’s body [4], a faster screening tool is necessary for COVID-19 outbreaks. As an alternative to the traditional PCR technique, researchers have proposed the use of radiography techniques like Computed Tomography (CT) scans and chest X-rays (CXRs) for COVID-19 screening. Early studies of COVID-19 positive patients have shown that their CT scans and CXRs show identifiable abnormalities [5], [6]. The idea is further strengthened by observing the high correlation between the PCR and radiological results, as demonstrated in    [7]. In [8] and [9], authors establish that the sensitivity of CT scan imaging outperforms the conventional PCR technique. The possible reasons may be immature development of nucleic acid detection technology, low patient viral load or improper clinical sampling, as stated in [8]. According to [6], [10], [11], and [12], the infestation of COVID-19 can primarily be characterized through radiographs by patches of ground-glass opacity (GGO) and consolidation (See Fig. 4). Additionally, authors in [4] and [13] have provided a deep insight into the statistical growth of radiological cues in COVID-19 positive patients and the temporal stages of the disease’s growth in a host’s body, respectively. These paved the way to a faster screening procedure than the PCR. Despite the radiological findings, there still exists a problem to use radiography as the primary screening tool for COVID-19. The problem being the lack of skilled radiologists across the globe. Ever since the advent of digital radiography, CT scans and CXRs have been used globally, but the final interpretations are required to be done by the experts, which could be time-consuming. Besides, authors in [5] have demonstrated through an experiment that sensitivity and specificity of screening COVID-19 positive CT images fluctuate significantly when done by radiologists. Therefore, for mass screening, automated or specifically AI-driven tools are necessary to be deployed across the globe, particularly in resource-constrained regions.

For all healthcare and/or (bio)medical problems, for more than a decade, deep learning has been a pinnacle in automation, especially in medical imaging. This motivates its use in the COVID-19 screening. Recently in [14], the author stated that to detect COVID-19, AI-driven tools are expected to have active-learning based cross-population train/test models that employ multitudinal and multimodal data. In this work [14], the use of deep learning and image data, such as CT scans and CXRs are addressed. Even though multimodal data can improve confidence in decision-making, for the COVID-19 case, much data are not available as of today.  Due to  lack of data, COVID-19 reveals limits of AI-driven tools. As soon as the COVID-19 pandemic came into play, several systems have been released to automate the screening procedure. Alibaba released an AI-based system to screen COVID-19 infection from CT scans, with an accuracy of 96% [15]. Researchers in [16] proposed a Convolutional Neural Network (CNN) based technique to differentiate COVID-19 from Pneumonia and normal cases of CT scans, with classification sensitivity of 0.90 for COVID-19. In [17], researchers have used a 3-dimensional deep learning model to segment infected regions from CT scans, followed by an attention driven network to classify COVID-19 from Influenza-A viral pneumonia and normal cases, and an accuracy of 86.7% was reported. Further, researchers in [18] have also proposed image segmentation schemes to detect lesions in CT scans. The prospective system is based on the popular UNet++ architecture, and it produces bounding boxes around lesion regions. The system achieved a  result of 100% per patient and 94.34% per image sensitivity. Authors in [19] and [20] have proposed deep learning models to classify COVID-19 positive CXRs from normal and Pneumonia cases, respectively. In [19], authors have investigated various standard CNN models, such as ResNet and AlexNet to extract deep features, which was followed by a Support Vector Machine (SVM) to classify COVID-19 positive cases. A maximum accuracy of 95.38% was reported using ResNet50 as the feature extractor. The latter work proposed a tailored CNN model using residual connections to achieve promising results of 80% positive predictive value (PPV) [20]. Additionally, authors in [21] have demonstrated the use of ResNet50, Inception Net V3 and Inception-Resnet V2 for identifying COVID-19 positive CXRs from healthy ones. arc On the whole, researchers found that the use of chest radiographs is better in terms of lung abnormalities screening [11], [14], [22]). With these, COVID-19 can be analyzed better using radiological image data [7], [8].

In this work, considering the fact that X-ray imaging systems are more prevalent and cheaper than CT scan systems, we use deep learning to screen COVID-19 using CXRs. We propose a CNN-based model, which we call Truncated Inception Net, solely based on the Inception Net  V3 architecture [23] (Fig. 1). The proposed model is validated to classify COVID-19 positive CXRs from Pneumonia, Tuberculosis and healthy CXRs. Using the limited number of COVID- 19 positive CXRs, made available by Cohen [24], our model demonstrates higher classification rate (99.96% accuracy with AUC of 1.0; Section III-C), where additional datasets are taken into account. As the COVID-19 dataset collection, alone, is not trivial, our experimental datasets that are composed of COVID-19, Pneumonia, Tuberculosis, and healthy cases, are sufficient to validate COVID-19 positive cases. For this, several publicly available datasets, such as Pneumonia dataset [25], Tuberculosis datasets (Shenzhen, China, and Montgomery County, USA) [26] are used to create six different experimental tests. On the whole, in this paper, we demonstrate that the Truncated Inception Net deep learning model outperforms the state-of-the-art results for COVID-19 positive cases.

The structure of the paper is summarized as follows. The proposed methodology and its various aspects of implementation are documented in Section II. In Section III, we provide datasets, validation protocol (including evaluation metrics). It also includes experimental results on six different datasets for COVID-19 positive cases. To prove the robustness and flexibility of our proposed method, a comparative study with respect to the state-of-the-art works is presented in Section IV. The paper is concluded in Section V.

2. Proposed Method

Inspired from the fact that COVID-19 shows patches of ground glass opacity (GGO) and consolidation in CXRs [10], to detect COVID-19 positive cases, a multi-resolution analysis of  the CXR images is deemed useful. The required trait is possessed by the Inception module, which is the fundamental block of the popular Inception Net V3 [23]. Additionally, considering the fact that the number of data samples of COVID-19 positive CXRs is very scarce at present, a modified version of the Inception Net V3 model [23] is proposed, which we call Truncated Inception Net. The Truncated Inception Net is primarily designed to avoid possible overfitting due to the lack of COVID-19 positive samples. Further, the Truncated Inception Net is computationally efficient. Originally, the Inception Net model was used on the ImageNet database, consisting of more  than 1.3 million images from 1000 different classes. In what follows, the different aspects of    the Truncated Inception Net model architecture and implementation are discussed.

A. Inception module

 The multi-resolution analysis capability of the Inception module comes from its inherent architecture. In traditional CNN models, kernels of specific receptive field sizes are used in specific layers to capture features through the use of convolution. On the contrary, in an inception module, kernels of various receptive field sizes (1  1, 3  3, and 5  5) are used in parallel to  extract features of varying sizes. The extracted parallel features are then stacked depth-wise to form the output of the inception module. A 3 3 max pooled version of the input is also stacked along with the previous feature maps. The combined output of the inception module provides rich feature maps of varying perspectives as inputs to the next convolutional layer of the CNN. Such a property of the Inception module explains its unique performance in medical imaging, and in our case, on the COVID-19 CXRs. For better understanding, the schematic representation of the Inception module is presented in Fig. 2.

B. Truncated architecture

Since the original Inception Net V3 model was built for the ImageNet database, the archi- tectural complexity of the model is well justified. On the contrary, the COVID-19 dataset used in our work is immensely small compared to the ImageNet database. Therefore, a truncation of the model is necessary to reduce the model complexity and eventually the number of trainable parameters to prevent the model from overfitting issue. The model was truncated at a point, where it retained 3 Inception modules and 1 grid size reduction block from the beginning, followed by the cascading of a Max Pooling and a Global Average  Pooling layer to reduce      the output dimension. The point of truncation was chosen experimentally, that yielded the best classification results. Finally, a fully connected layer was cascaded to perform the classification task. The truncation of the model not only reduced training time and trainable parameters but also reduced the processing time while evaluating a CXR to detect COVID-19 positive cases.   As a consequence, it facilitates mass screening at an efficient speed and accuracy. For more detailed information about computational efficiency, we refer to Section III. The architecture of the complete Truncated Inception Net can be visualized in Fig.

C. Adaptive learning rate protocol

For training the Truncated Inception Net, a dynamic protocol was used to control the learning rate at each epoch because of the following reasons: a) A constant learning rate of high value often leads to divergence of the weight vectors’ trajectory from global minimum in the loss function space; and b) an arbitrarily chosen low value often takes longer period of training time. Therefore, a dynamic procedure was opted, where the learning rate was initialized with a value  of 0.001 and then it was reduced by a factor of 2 every time the validation loss remained same   or did not decrease for more than 3 epochs. In this case, the factor of 3 epochs is known as the patience factor. The process is well explained through the diagram in Fig. 3. This procedure yielded the behavior of reducing the velocity of approach when the weight vectors are close to the global minimum, to prevent overshooting.

D. Transfer learning

Deep learning models are inherently data intensive. However,  since the size of the COVID-  19 dataset is very small compared to standard datasets used in deep learning, the concept of transfer learning can be applied to augment the decision-making process. Transfer learning uses the concept of transferring knowledge from one domain to another by using trained weights  from the previous domain. Traditionally in a CNN, the weight matrices of several layers from  the beginning are frozen while training on the secondary domain, and only the remaining layers are fine-tuned. This process works well when both the domains share an overlapping region in the low-level features. In our case, since the ImageNet and the COVID-19 datasets belong to non-overlapping domains, the trained weights from the ImageNet dataset were used to initialize the weights of our model, but none of them were frozen. This kept all the layers initialized with relatively more meaningful weights than random initialization, and subject to learning during  the training procedure.

3. Data And Experiments

A. Datasets

Collecting COVID-19 dataset is not trivial. We, however, collect a number of CXR benchmark collections (C1 to C3) from the literature (See Table I). They help to showcase/validate the usability and robustness of our model.

C1: COVID-19 collection [24] is an open-source collection that is made available and main- tained by Joseph Paul Cohen. As of now, it is composed of 73 COVID-19 positive CXRs, along with some other CXRs of diseases like MERS, SARS, and viral Pneumonia. For   our purpose, only COVID-19 positive posteroanterior CXRs are considered.

C2: Pneumonia collection [25] (Kaggle CXR collection) is composed of 5863 CXRs. Out of this, 1583 CXRs are normal or healthy CXRs and the remaining 4280 CXRs show various manifestations of viral and bacterial Pneumonia.

C3: Two publicly available Tuberculosis (TB) collections [26] are considered: a) Shenzhen, China and b) Montgomery County, USA. These CXR benchmark collections were made available by the U.S. National Library of Medicine, National Institutes of Health (NIH). The Shenzhen, China collection is composed of 340 normal cases and 342 positive cases  of TB. The Montgomery County,  USA collection is composed of 80 normal CXRs and    58 TB positive CXRs.

Few samples from the aforementioned collections are visualized in Fig. 4. Using aforemen- tioned collections, we constructed six different combinations of data to train and validate our model. As provided in Table II, these six different combinations of datasets (D1 to D6) are enlisted below:

D1: In dataset D1, 73 COVID-19 positive CXRs and 340 healthy CXRs from the Shenzhen, China collections are considered.

D2: For this dataset D2, 73 COVID-19 positive CXRs and 80 healthy CXRs from the Mont- gomery County, USA are considered.

D3: D3 consists of 73 COVID-19 positive CXRs and 1583 healthy CXRs from the Pneumonia collections are considered.

TABLE I

DATA COLLECTION (PUBLICLY AVAILABLE).

TABLE II

EXPERIMENTAL DATASETS  USING TABLE I.

Index: +ve = positive cases and -ve = negative/healthy cases.

D2: D4 contains 73 COVID-19 positive CXRs and 2003 healthy CXRs, combined from the Shenzhen, Montgomery and Pneumonia collections are considered.

D5: In dataset D5, 73 COVID-19 positive CXRs, 4280 Pneumonia positive CXRs and 1583 healthy CXRs from the Pneumonia collections are considered.

D6: In dataset D6, 73 COVID-19 positive CXRs and 6683 non-COVID CXRs (comprising of 4280 Pneumonia positive, 400 TB positive and 2003 healthy CXRs) are considered.

The primary motivation behind constructing the various data combinations (D1 to D6) is to show the robustness of the Truncated Inception Net to detect COVID-19 positive cases. Further, COVID-19 is believed to have a close relationship with traditional Pneumonia. Therefore, a separate dataset (D5) was constructed to show whether our proposed model is able to differentiate COVID-19 positive cases from those traditional Pneumonia positive cases. Besides, CXRs of Tuberculosis manifestation were also added in D6 to prove that our model is robust enough to identify COVID-19 from other diseases like TB, Pneumonia, and healthy CXRs. The robustness also lies in the way we collect data, where regional variation can be considered as a crucial element. In our datasets, the healthy CXRs in D1, D2, and D3 are collected from different regions of the world. Considering multiple combination of data from different places can help develop cross-population train/test models¹.

¹Even though, our tests proved that the proposed model can be considered as a cross-population train/test model, it is beyond the scope of the paper.

As an input to our model, CXR images were scaled down to the size of 224   224   3 to    match the input dimensions of the Truncated Inception Net. Such a resizing can also reduce computational complexity. Further, the images were normalized using the min-max scaling scheme.

B. Validation protocol and evaluation metrics

To  validate our proposed model, a 10-fold cross-validation scheme was opted for training   and testing purposes on all six datasets: D1 – D6. A cross-validation scheme ensures that the model’s performance is not biased by the presence of outlier data samples in the training or testing datasets. For each of the 10-folds, six different evaluation metrics were employed: a) Accuracy (ACC); b) Area under the ROC curve (AUC); c) Sensitivity (SEN); d) Specificity (SPEC); e) Precision (PREC); and f) F1 score. These can be computed as follows:

ACC = (t_p + t_n)/(t_p + t_n + f_p + f_n), SEN = t_p/(t_p + f_n),

SPEC = t_n/(t_n + f_p), PREC = t_p/(t_p + f_p), and F1 score = 2 ((PREC × SEN)/(PREC + SEN)) ,

where t_p, f_p, t_n, and f_n are the total number of true positives, false positives, true negatives, and false negatives. The mean scores from all 10 folds were taken for each of the above metrics, to get the final results on a particular dataset.

In traditional deep learning tasks, a primary metric like accuracy is sufficient to judge the performance of a deep learning model. On the contrary, such an assumption does not work well when considering imbalanced datasets. In such cases (more often in medical dataset), the positive class to be predicted often has much lower data samples than the negative class. Therefore, accuracy would demonstrate a fairly high value even if the model labels all the test data to be negative. Therefore, special attention is given to metrics like Sensitivity/Recall, Precision, and  F1 score here.

In the context of COVID-19, the SEN metric plays a very crucial role when deploying a model for screening patients in the early stages of a pandemic. Sensitivity measures the likelihood that the model would not miss to classify COVID-19 positive samples/patients. This prevents the further spreading of the infection. Secondly, the precision measures the likelihood that a model would not make a mistake to classify normal patients as COVID-19 positive. This metric becomes very important in the later stages of a pandemic, when medical resources are limited, and they are available only to the patients that are in need. Besides, F1 score is used to extract the combined performance score of a model, which is basically the harmonic mean of the precision and sensitivity of a model.

C. Results and analysis

Before providing quantitative results, we first provide activation maps generated by our pro- posed model for a COVID-19 positive, Pneumonia positive, and TB positive CXR can be visualized in Fig. 5. Qualitatively speaking, such feature maps help understand how different their features could possibly be.

Following the validation protocol and evaluation metrics mentioned in the previous Section III-B, we present the mean scores that were achieved using 10-fold cross validation train-test scheme, on each of the six different datasets: D1 – D6. The experimental results are well documented in Table III. Also, standard deviation(σ) is reported all cases, whose very low value proves the statistical robustness of our model. Our proposed Truncated Inception Net model achieves a classification ACC, AUC, SEN, SPEC, PREC, and F1 score of 99.96%, 1.0, 0.98, 0.99, 0.98, and 0.98, respectively, on the dataset: D5 (COVID-19 positive case detection against Pneumonia and healthy cases) and that of 99.92%, 0.99, 0.93, 1.0, 1.0, and 0.96, respectively, on the D6 dataset (COVID-19 positive case detection against Pneumonia, TB, and healthy CXRs). Since the custom datasets being used were highly imbalanced in terms of class representation, sensitivity and precision are the most significant metrics in our case, as said in Section III-B. Consequently, the proposed model achieves high sensitivity and precision on these datasets. For a better understanding of the results, six different ROC curves are shown in Fig. 6; one for each dataset, starting from D1 to D6.

TABLE III

RESULTS: AVERAGE ACC IN %, AUC, SEN, SPEC, PREC, AND F1 SCORE USING 10 FOLD CROSS-VALIDATION WITH σ

STANDARD DEVIATION.

Additionally, since for every dataset we computed 10-fold cross validation, for better under- standing of how average scores and their standard deviation were computed, the results obtained from each fold on the dataset: D6 are provided in Table IV. Besides, the proposed Truncated Inception Net model performs 2.3 0.18 times on an average faster than Inception Net V3  model. In Table  V, computational times (by taking 10 different CXR samples as input) are    used to demonstrate the differences between them. The primary reason being the large number  of parameters in the original Inception Net V3 model. Precisely, this model contains more than 21.7 million trainable parameters in contrast our model which contains only 2.1 million trainable parameters, making it a better choice for training on small datasets and also for active learning. Therefore, for mass screening in resource-constrained areas, employing a faster tool is the

4. Comparative Study

Since COVID-19 outbreak, very few pieces of works have been proposed/reported using CXRs to detect COVID-19 positive cases (see Section I): In our comparison, ResNet50 and SVM [19], COVID-Net [20], ResNet50 [21], and Inception Net V3 [21] are considered even though they are not peer-reviewed research articles. We have compared with these pieces works using exact same evaluation metrics (ACC in %, AUC, SEN, SPEC, PREC, and F1 score) and nature of dataset.

TABLE IV

RESULTS: ACC IN %, AUC, SEN, SPEC, PREC, AND F1 SCORE FOR EACH FOLD OF 10 FOLD CROSS-VALIDATION ON THE

D6 DATASET.

TABLE V

COMPARISON: COMPUTATIONAL TIME (IN MS) BETWEEN INCEPTION NET V3 (FULL ARCHITECTURE) AND TRUNCATED

INCEPTION NET.

Like other works, we take COVID-19 positive and healthy CXRs from Pneumonia dataset (D3 in our case), and used this result as a comparison to other works. Besides, since all models were based on deep learning models, we consider an essential element i.e., number of parameters in our comparison. Table VI provides a complete comparative study. Not all the authors reported AUC, SPEC, and F1 score. On the whole, considering the number of parameters, the proposed Truncated Inception Net outperforms all. Note that, since our model is the derivative of Inception Net V3 model, it is worth to compare between them. We observe that, in both computational  time (Table V) and performance scores (Table  VI), Truncated Inception Net performs better  than Inception Net V3 [21]. This suggests that the Truncated Inception Net is not only more computationally effective in terms of training and usability, but also more flexible for the purpose of active learning [14].

5. Conclusion And Future Works

In this work, we have proposed the Truncated Inception Net deep learning model to screen/detect COVID-19 positive patients using chest X-rays. For validation, experimental tests were done on six different experimental datasets by combining COVID-19 positive, Pneumonia positive, Tuberculosis positive, and healthy CXRs. The proposed model outperforms the state-of-the-art results in detecting COVID-19 cases from non-COVID ones. Besides, considering the number of parameters used in our proposed model, it is computationally efficient as compared to original Inception Net V3 model and other works proposed in the literature.

TABLE VI

COMPARISON TABLE

†Not peer-reviewed research articles (source: https://arxiv.org)

Observing the performance scores, the Truncated Inception Net can serve as a milestone for screening COVID-19 under active-learning framework on multitudinal/multimodal data [14]. It also motivates to work on cross-population train/test models. Integrating this model with CheXnet model [27] will be our immediate plan, since ChexNet is primarily employed to analyze CXRs.

Declarations

Conflict of interest: Authors declared no conflict of interest.

Ethical approval: This article does not contain any studies with human participants performed  by any of the authors.

References

1. World  Health Naming the coronavirus disease (covid-19) and the virus that causes it. https://www.who.int/emergencies/diseases/novel-coronavirus-2019/technical-guidance/naming-the-coronavirus-disease- (covid-2019)-and-the-virus-that-causes-it, 2020.
2. World Health Organization. The continuing 2019-ncov epidemic threat of novel coronaviruses to global health - the latest 2019 novel coronavirus outbreak in wuhan, https://pubmed.ncbi.nlm.nih.gov/31953166/, 2020.
3. World Health Coronavirus disease (covid-2019) situation reports. https://www.who.int/emergencies/diseases/novel-coronavirus-2019/situation-reports, 2020.
4. Mingzhi Li, Pinggui Lei, Bingliang Zeng, Zongliang Li, Peng Yu, Bing Fan, Chuanhong Wang, Zicong Li, Jian Zhou, Shaobo Hu, et al. Coronavirus disease (covid-19): Spectrum of ct findings and temporal progression of the disease. Academic Radiology,
5. Harrison X Bai, Ben Hsieh, Zeng Xiong, Kasey Halsey, Ji Whae Choi, Thi My Linh Tran, Ian Pan, Lin-Bo Shi, Dong-Cui Wang, Ji Mei, et al. Performance of radiologists in differentiating covid-19 from viral pneumonia on chest ct. Radiology, page 200823,
6. Alexander Gross,  Dorina  Thiemig,  Franz-Wilhelm  Koch,  Martin  Schwarz,  Sven  Gla¨ser,  and  Thomas      Ct appearance of severe, laboratory-proven coronavirus disease 2019 (covid-19) in a caucasian patient in berlin, germany. In Ro¨Fo-Fortschritte auf dem Gebiet der Ro¨ntgenstrahlen und der bildgebenden Verfahren. © Georg Thieme Verlag KG, 2020.
7. Tao Ai, Zhenlu Yang, Hongyan Hou, Chenao Zhan, Chong Chen, Wenzhi Lv, Qian Tao, Ziyong Sun, and Liming Xia. Correlation of chest ct and rt-pcr testing in coronavirus disease 2019 (covid-19) in china: a report of 1014 cases. Radiology, page 200642,
8. Yicheng Fang, Huangqi Zhang, Jicheng Xie, Minjie Lin, Lingjun Ying, Peipei Pang, and Wenbin   Sensitivity of chest   ct for covid-19: comparison to rt-pcr. Radiology, page 200432, 2020.
9. Chung JH Elicker BM Ketai LH Kanne JP, Little Essentials for radiologists on covid-19: An update-radiology scientific expert panel. Radiology, page 200527, 2020 Feb 27.
10. Weifang Kong and Prachi P Agarwal. Chest imaging appearance of covid-19 infection. Radiology: Cardiothoracic Imaging, 2(1):e200028,
11. Chaolin Huang, Yeming Wang, Xingwang Li, Lili Ren, Jianping Zhao, Yi Hu, Li Zhang, Guohui Fan, Jiuyang Xu, Xiaoying Gu, et al. Clinical features of patients infected with 2019 novel coronavirus in wuhan, china. The Lancet, 395(10223):497–506,
12. Ming-Yen Ng, Elaine YP Lee, Jin Yang, Fangfang Yang, Xia Li, Hongxia Wang, Macy Mei-sze Lui, Christine Shing-Yen Lo, Barry Leung, Pek-Lan Khong, et al. Imaging profile of the covid-19 infection: radiologic findings and literature Radiology: Cardiothoracic Imaging, 2(1):e200034, 2020.
13. Yuhui Wang, Chengjun Dong, Yue Hu, Chungao Li, Qianqian Ren, Xin Zhang, Heshui Shi, and Min Zhou. Temporal changes of ct findings in 90 patients with covid-19 pneumonia: A longitudinal Radiology, page 200843, 2020.
14. KC Santosh. Ai-driven tools for coronavirus outbreak: Need of active learning and cross-population train/test models on multitudinal/multimodal Journal of Medical Systems, 44(5):1–5, 2020.
15. Technology org,   ai   algorithm   detects   coronavirus   infections   in   patients   from   ct   scans   with   96%   accu-   https://www.technology.org/2020/03/01/ai-algorithm-detects-coronavirus-infections-in-patients-from-ct-scans-with-96- accuracy, 2020(accessed March 02, 2020).
16. Lin Li, Lixin Qin, Zeguo Xu, Youbing Yin, Xin Wang, Bin Kong, Junjie Bai, Yi Lu, Zhenghan Fang, Qi Song, et al. Artificial intelligence distinguishes covid-19 from community acquired pneumonia on chest ct. Radiology, page 200905,
17. Xiaowei Xu, Xiangao Jiang, Chunlian Ma, Peng Du, Xukun Li, Shuangzhi Lv, Liang Yu, Yanfei Chen, Junwei Su, Guanjing Lang, et al. Deep learning system to screen coronavirus disease 2019 pneumonia. arXiv preprint arXiv:2002.09334,
18. Jun Chen, Lianlian Wu, Jun Zhang, Liang Zhang, Dexin Gong, Yilin Zhao, Shan Hu, Yonggui Wang, Xiao Hu, Biqing Zheng, et al. Deep learning-based model for detecting 2019 novel coronavirus pneumonia on high-resolution computed tomography: a prospective study. medRxiv,
19. Prabira Kumar Sethy and Santi Kumari Detection of coronavirus disease (covid-19) based on deep features. Preprints, 2020.
20. Linda Wang and Alexander Covid-net: A tailored deep convolutional neural network design for detection of covid-19 cases from chest radiography images. https://arxiv.org/abs/2003.09871, 2020.
21. Ali Narin, Ceren Kaya, and Ziynet Pamuk. Automatic detection of coronavirus disease (covid-19) using x-ray images and deep convolutional neural arXiv preprint arXiv:2003.10849, 2020.
22. KC Santosh and Sameer Antani. Automated chest x-ray screening: Can lung region symmetry help detect pulmonary abnormalities? IEEE transactions on medical imaging, 37(5):1168–1177,
23. Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016.
24. Joseph Paul Cohen. Covid-19 image data collection. https://github.com/ieee8023/covid-chestxray-dataset,
25. Paul Kaggle chest x-ray images (pneumonia) dataset. https://www.kaggle.com/paultimothymooney/chest-xray- pneumonia/, 2020.
26. S. National Library of Medicine. Tuberculosis chest x-ray image data sets. https://ceb.nlm.nih.gov/tuberculosis-chest-x- ray-image-data-sets/, 2020.
27. Pranav Rajpurkar, Jeremy Irvin, Kaylie Zhu, Brandon Yang, Hershel Mehta, Tony Duan, Daisy Ding, Aarti Bagul, Curtis Langlotz, Katie Shpanskaya, et al. Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225,