COVID-19 infection map generation and detection from chest X-ray images

Computer-aided diagnosis has become a necessity for accurate and immediate coronavirus disease 2019 (COVID-19) detection to aid treatment and prevent the spread of the virus. Numerous studies have proposed to use Deep Learning techniques for COVID-19 diagnosis. However, they have used very limited chest X-ray (CXR) image repositories for evaluation with a small number, a few hundreds, of COVID-19 samples. Moreover, these methods can neither localize nor grade the severity of COVID-19 infection. For this purpose, recent studies proposed to explore the activation maps of deep networks. However, they remain inaccurate for localizing the actual infestation making them unreliable for clinical use. This study proposes a novel method for the joint localization, severity grading, and detection of COVID-19 from CXR images by generating the so-called infection maps. To accomplish this, we have compiled the largest dataset with 119,316 CXR images including 2951 COVID-19 samples, where the annotation of the ground-truth segmentation masks is performed on CXRs by a novel collaborative human–machine approach. Furthermore, we publicly release the first CXR dataset with the ground-truth segmentation masks of the COVID-19 infected regions. A detailed set of experiments show that state-of-the-art segmentation networks can learn to localize COVID-19 infection with an F1-score of 83.20%, which is significantly superior to the activation maps created by the previous methods. Finally, the proposed approach achieved a COVID-19 detection performance with 94.96% sensitivity and 99.88% specificity.


Introduction
Coronavirus disease 2019 (COVID-19) caused by severe acute respiratory syndrome Coronavirus-2 (SARS-CoV-2) was first reported in December 2019 in Wuhan, China. The highly infectious disease rapidly spread around the World with millions of positive cases. As a result, COVID-19 was declared as a pandemic by the World Health Organization in March 2020. The disease may lead to hospitalization, intubation, intensive care, and even death, especially for the elderly [1,2]. Naturally, reliable detection of the disease has the utmost importance. However, the diagnosis of COVID-19 is not straight-forward since its symptoms, such as cough, fever, breathlessness, and diarrhea are generally indistinguishable from other viral infections [3,4].
The diagnostic tools to detect COVID-19 are currently reverse transcription of polymerase chain reaction (RT-PCR) assays and chest imaging techniques, such as Computed Tomography (CT) and X-ray imaging. Primarily, RT-PCR has become the gold standard in the diagnosis of COVID-19 [5,6]. However, RT-PCR arrays have a high false alarm rate which may be caused by the virus mutations in the SARS-CoV-2 genome, sample contamination, or damage to the sample acquired from the patient [7,8]. In fact, it is shown in hospitalized patients that RT-PCR sensitivity is low and the test results are highly unstable [6,[9][10][11]. Therefore, it is recommended to perform chest CT imaging initially on the suspected COVID-19 cases [12], since it is a more reliable clinical tool in the diagnosis with higher sensitivity compared to RT-PCR. Hence, several studies [12][13][14] suggest performing CT on the negative RT-PCR findings of the suspected cases. However, there are several limitations of CT scans. Their sensitivity is limited in the early COVID-19 phase groups [15], and they are limited to recognize only specific viruses [16], slow in image acquisition, and costly. On the other hand, X-ray imaging is faster, cheaper, and less harmful to the body in terms of radiation exposure compared to CT [17,18]. Moreover, unlike CT devices, X-ray devices are easily accessible; hence, reducing the risk of COVID-19 contamination during the imaging process [19]. Currently, chest X-ray (CXR) imaging is widely used as an assistive tool in COVID-19 prognosis, and it is reported to have a potential diagnosis capability in recent studies [20].
In order to automate COVID-19 detection/recognition from CXR images, several studies [21][22][23] have extracted features from the CXRs to utilize Support Vector Machines classifier. On the other hand, many studies [17,[24][25][26][27][28][29][30][31][32] have proposed to use deep Convolutional Neural Networks (CNNs). However, the main limitation of these studies is that the data is scarce for the target COVID-19 class. Such a limited amount of data degrades the learning performance of the deep networks. Two recent studies [33] and [34] have addressed this drawback with a compact network structure and achieved the stateof-the-art detection performance over the benchmark QaTa-COV19 and Early-QaTa-COV19 datasets that consist of 462 and 175 COVID-19 CXR images, respectively. Although these datasets were the largest available at that time, such a limited number of COVID-19 samples raises robustness and reliability issues for the proposed methods in general.
Moreover, all these previous machine learning solutions with X-ray imaging remain limited to only COVID-19 detection. However, as stated by Shi [35], COVID-19 pneumonia screening is important for evaluating the status of the patient and treatment. Therefore, along with the detection, COVID-19 related infection localization is another crucial problem. Hence, several studies [36][37][38] produced activation maps that are generated from different Deep Learning (DL) models trained for COVID-19 detection (classification) task to localize COVID-19 infection in the lungs. Infection localization has two vital objectives: an accurate assessment of the infection location and the severity of the disease. However, the results of previous studies show that the activation maps generated inherently from the underlying DL network may fail to accomplish both objectives, that is, irrelevant locations with biased severity grading appeared in many cases. To overcome these problems, two studies [39,40] proposed to perform lung segmentation as the first step in their approaches. This way, they have narrowed the region of interest down to the regions of lungs to increase the reliability of their methods. Overall, until this study, screening COVID-19 infection from such activation maps produced by classification networks was the only option for the localization due to the absence of ground-truth of the datasets available in the literature. Many studies [35,39,[41][42][43] have COVID-19 infection ground-truths for CT images; however, ground-truth segmentation masks for CXR images are non-existent.
In this study, in order to overcome the aforementioned limitations and drawbacks, first, the benchmark dataset QaTa-COV19 proposed by the researchers of Qatar University and Tampere University in [33] and [34] is extended to include 2951 COVID-19 samples. This new dataset is 3-20 times larger than those used in earlier studies. The extended benchmark dataset, QaTa-COV19 with around 120K CXR images, is not only the largest ever composed dataset, but it is the first dataset that has the ground-truth segmentation masks for COVID-19 infection regions, as some samples are shown in Fig. 1. A crucial property of QaTa-COV19 dataset is that it contains CXRs with other (non-COVID- 19) infections and anomalies such as pneumonia and pulmonary edema, Fig. 1 The COVID-19 sample CXR images, their corresponding ground-truth segmentation masks which are annotated by the collaborative human-machine approach, and the generated infection maps from the state-of-the-art segmentation models Degerli et al. Health Inf Sci Syst (2021) 9 :15 both of which exhibit high visual similarity to COVID-19 infection in the lungs. Therefore, this is significantly more challenging task than distinguishing COVID-19 from the normal (healthy) cases as almost all studies in the literature did.
To obtain the ground-truth segmentation masks for the COVID-19 infected regions, a human-machine collaborative approach is introduced. The objective is to significantly reduce the human labor and thus to speed up and also to improve the segmentation masks because when they are drawn solely by medical doctors (MDs), human error due to limited perception, hand-crafting, and subjectivity will deteriorate the overall quality. This is an iterative process, where MDs initiate the segmentation by "manually-drawn" segmentation masks for a subset of CXR images. Then, the trained segmentation networks over this subset generate their own "competing" masks and the MDs are asked to compare them pair-wise (initial manual segmentation vs. machine-segmented masks) for each patient. Such a verification improves the quality of the generated masks as well as the (following) training runs. Over the best masks selected by experts, the networks are trained again this time over a larger set (or even perhaps over the entire dataset), and among the masks generated by the networks, the best masks are selected by the MDs. This human-machine collaboration process continues until the MDs are fully satisfied, i.e., a satisfactory mask can be found among the masks generated by the networks for all CXR images in the dataset. In this study, we show that even with two stages (iterations), highly superior infection maps can be obtained using which an elegant COVID-19 detection performance can be achieved.
The rest of the paper is organized as follows. In "The benchmark QaTa-COV19 dataset", we introduce the benchmark QaTa-COV19 1 dataset. Our novel human-machine collaborative approach for the ground-truth annotation is explained in "Collaborative human-machine ground-truth annotation". Next, the details of COVID-19 infected region segmentation, and the infection map generation and COVID-19 detection are presented in "COVID-19 infected region segmentation" and "Infection map generation and COVID-19 detection", respectively 2 . The experimental setup and results with the benchmark dataset are reported in "Experimental setup" and "Experimental results", respectively. Finally, we conclude the paper in "Conclusions".

Materials and methodology
The proposed approach in this study is composed of three main phases: (1) training the state-of-the-art deep models for COVID-19 infected region segmentation using the ground-truth segmentation masks, (2) infection map generation from the trained segmentation networks, and (3) COVID-19 detection as it can be depicted in Fig. 2. In this section, we first detail the creation of the benchmark QaTa-COV19 dataset. Then, the proposed approach for collaborative human-machine ground-truth generation is introduced.

The benchmark QaTa-COV19 dataset
The researchers of Qatar University and Tampere University have compiled the largest COVID-19 dataset up to date with nearly 120K CXR images: QaTa-COV19 including 2951 COVID-19 CXRs. To create QaTa-COV19, we have utilized several publicly available, scattered, and different format datasets and repositories. Therefore, the collected images from the datasets had some duplicate, over-exposed and low-quality images that were identified and removed in the preprocessing stage. Consequently, the COVID-19 CXRs are from different publicly available sources resulting in high intra-class dissimilarity as depicted in Fig. 3. The image sources of COVID-19 and control group CXRs are detailed as follows:

COVID-19 CXRs
BIMCV-COVID19+ [44] is the largest publicly available dataset with 2473 COVID-19 positive CXR images. The CXR images of BIMCV-COVID19+ dataset were recorded with computed radiography (CR) and digital X-ray (DX) machines. Hannover Medical School and Institute for Diagnostic and Interventional Radiology [45] released 183 CXR images of COVID-19 patients. A total of 959 CXR images are from public repositories: Italian Society of Medical and Interventional Radiology (SIRM), GitHub, and Kaggle [40,[46][47][48][49]. As mentioned earlier, any duplication and low-quality images are removed since COVID-19 CXR images are collected from different public datasets and repositories. In this study, a total of 2951 COVID-19 CXRs are gathered from the aforementioned datasets. Therefore, COVID-19 CXRs are of different age, group, gender, and ethnicity.

Control group CXRs
In this study, we have considered two control groups in the experimental evaluation. Group-I consists of only normal (healthy) CXRs with a smaller number of images compared to the second group. RSNA pneumonia detection challenge dataset [50] is comprised of about 29.7K CXR images, where 8851 images are normal. All CXRs in the dataset are in DICOM format, a popularly used format for medical imaging. Padchest dataset [51] consists of 160, 868 CXR images from 67, 625 patients, where 37, 871 images are from normal class. The images are evaluated and reported by radiologists at Hospital Sun Juan in Spain during 2009-2017. The dataset includes six different position views of CXR and additional information regarding image acquisition and patient demography. Paul Mooney [52] has released an X-ray dataset of 5863 CXR images from a total of 5856 patients, where 1583 images are from normal class. The data is collected from pediatric patients aging one to five years old at Guangzhou Women and Children's Medical Center, Guangzhou. The dataset in [53] consists of 7470 CXR images and the corresponding radiologist reports from the Indiana Network for Patient Care, where a total of 1343 frontal CXR samples are labeled as normal. In [54], there are 80 normal CXRs from the tuberculosis control program of the Department of Health and Human Services of Montgomery County and 326 normal CXRs from Shenzhen Hospital. In this study, a total of 12, 544 normal CXRs are included in control Group-I from the aforementioned datasets. On the other hand, Group-II consists of 116, 365 CXRs from 15 different classes. ChestX-ray14 [55] consists of 112, 120 CXRs with normal and 14 different thoracic disease images, which are atelectasis, cardiomegaly, effusion, infiltration, mass, nodule, pneumonia, pneumothorax, consolidation, edema, emphysema, fibrosis, PT, hernia, and normal (no findings). Additionally, from the pediatric patients [52], 2760 bacterial and 1485 viral pneumonia CXRs are included in Group-II.

Collaborative human-machine ground-truth annotation
Recent developments in the machine and deep learning techniques led to state-of-the-art performance in many computer vision (CV) tasks, such as image classification, object detection, and image segmentation. However, supervised DL methods require a huge amount of annotated data. Otherwise, the limited amount of data degrades the performance of the deep network structures since their generalization capability depends on the availability of large datasets. Nevertheless, to produce ground-truth segmentation masks, pixel-accurate image segmentation by human experts can be a cumbersome and highly subjective task even for moderate size datasets.
In order to overcome this challenge, in this study, we propose a novel collaborative human-machine approach to accurately produce the ground-truth segmentation masks for infected regions directly from the CXR images. The proposed approach is performed in two main stages. First, a group of expert MDs manually segment the infected regions of a subset of (500 in our case) CXR images. Then, several segmentation networks that are inspired by the U-Net [56] structure with a 5-fold cross-validation scheme, are trained over the initial ground-truth masks. For each fold, the segmentation masks of the test samples are predicted by the networks. The network predicted masks along with the initial (MD drawn) ground-truth masks, and original CXR image are assessed by the MDs, and the best segmentation mask among them is selected. Steps of Stage-I are illustrated in Fig. 4 (top). At the end of the first stage, collaboratively annotated ground-truth masks for the subset of CXR images are formed, and they are obviously superior to the initial manually drawn masks since they are selected by the MDs. An interesting observation in this stage was that MDs preferred the machine-generated masks over the manually drawn masks in the first stage in three out of five cases.
In the second stage five deep networks, inspired by U-Net [56], UNet++ [57], and DLA [58] architectures are trained over the collaborative masks, which were formed in Stage-I. The trained segmentation networks are used to predict the segmentation masks of the rest of the data, which is around 2400 unannotated COVID-19 images. Among the five predictions, the expert MDs select the best one as the ground-truth or deny all if none was found successful. For the latter case, MDs were asked to draw the ground-truth masks manually. However, we notice that this was indeed a minority case that included less than 5% of unannotated data. The steps of Stage-II are shown in Fig. 4 (bottom). As a result, the groundtruth masks for 2951 COVID-19 CXR images are gathered to construct the benchmark QaTa-COV19 dataset. The proposed approach does not only save valuable human labor time, but it also improves the quality and reliability of the masks by reducing the subjectivity with Stage-II verification step.

COVID-19 infected region segmentation
Segmentation of COVID-19 infection is the first step of our proposed approach as depicted in Fig. 2. Once the ground-truth annotation for QaTa-COV19 benchmark dataset is formed as explained in the previous section, we perform infected region segmentation extensively with 24 different network configurations. We have used three different segmentation models: U-Net, UNet++ and DLA, with four different encoder structures: CheXNet, DenseNet-121, Inception-v3 and ResNet-50, and frozen & not frozen encoder weight configurations.

Segmentation models
We have tried distinct segmentation model structures starting from shallow to deep structures with varied configurations as follows: -U-Net [56] is an outperforming network for medical image segmentation applications with a u-shaped architecture as the encoder part is symmetric with respect to its decoder part. Therefore, this unique decoder structure with many feature channels allows the network to carry the information through its latest layers. -UNet++ [57] has further developed the decoder structure of U-Net by connecting the encoder to the decoder with the nested dense convolutional blocks. This way, the bridge between the encoder and decoder parts are more firmly knit; thus, the information can be transferred to its final layers more intensively compared to the classic U-Net. -DLA [58] investigates the connecting bridges between the encoder and decoder, and proposes a way to fuse the semantic and spatial information with dense layers, which are progressively aggregated by iterative merging to deeper and larger scales.

Encoder selections for segmentation models
In this study, we use several deep CNNs to form the encoder part of the above-mentioned segmentation models as follows: -DenseNet-121 [59] is a deep network with 121 layers, each with additional input nodes connecting all the layers directly with each other. Therefore, the maximum information flow through the network is satisfied. -CheXNet [60] is based on the architecture of DenseNet-121, which is trained over the ChestX-ray14 dataset [55] to detect pneumonia cases from CXR images. In [60], DenseNet-121 is initialized with the ImageNet weights and fine-tuned over 100K CXR images resulting from the state-of-the-art results on the ChestX-ray14 dataset with a better performance compared to the conclusions of radiologists. -Inception-v3 [61] achieves state-of-the-art results with much less computational complexity compared to its deep competitors by factorizing the convolutions and pruning the dimensions inside the network. Despite the less complexity, it preserves a higher performance. -ResNet-50 [62] introduces a deep residual learning framework that forces the desired mapping of the input to a residual mapping. It is possible to achieve this goal by the shortcut connections on the stacked layers. These connections enable to merge the input and output of the stacked layers by addition operations; therefore, the problem of gradient vanishing is prevented.
We perform transfer learning on the encoder side of the segmentation models by initializing the layers with the ImageNet weights, except for CheXNet which is pretrained on the ChestX-ray14 dataset. We tried two configurations, in the first we freeze the encoder layers while in the second, they are allowed to vary.

Hybrid loss function
In this study, we have performed training the segmentation networks with a hybrid loss function by combining focal loss [63] with dice loss [64] to achieve a better segmentation performance. We use focal loss since COVID-19 infected region segmentation is an imbalanced problem: the number of background pixels is superior to the foreground's. Let the ground-truth segmentation mask be Y , where each pixel class label is defined as y, and the network prediction as ŷ . We define the pixel class probabilities as for the positive class P(y = 1) = p , and for the negative class P(y = 0) = 1 − p . On the other hand, the network prediction probabilities are modeled by the logistic function using the sigmoid curve as, where z is some function of the input CXR image X . Then, we define the cross-entropy (CE) loss as follows: A common solution to address the class imbalance problem is to add a weighting factor α ∈ [0, 1] for the positive class, and 1 − α for the negative class, which defines the balanced cross-entropy (BCE) loss as, Degerli et al. Health Inf Sci Syst (2021) 9:15 In this way, the importance of positive and negative samples are balanced. However, adding the α factor does not solve the issue for the large class imbalance scenario. This is because the network cannot distinguish outliers (hard samples) and inliers (easy samples) with the BCE loss. To overcome this drawback, focal loss [63] proposes to set focusing parameter γ ≥ 0 in order to down-weight the loss of easy samples that occur with small errors; so that the model can be forced to learn hard negative samples. The focal (F) loss is defined as, where F loss is equivalent to BCE loss when γ = 0 . In our experimental setup, we use the default setting as α = 0.25 , and γ = 2 for all the networks. To achieve a good segmentation performance, we combined focal loss with dice loss, which is based on the dice coefficient (DC) defined as follows: where Ŷ is the predicted segmentation mask of the network. Hence, the DC can be interpreted as a dice (D) loss as follows: where h and w are the height and width of the groundtruth and prediction masks Y and Ŷ , respectively. Finally, we combined D and F losses by summation to achieve the so-called hybrid loss function for the segmentation networks.

Infection map generation and COVID-19 detection
Having the training set of COVID-19 CXR images via the collaborative human-machine approach explained in "The benchmark QaTa-COV19 dataset", we train the aforementioned segmentation networks to produce infection maps. After training the segmentation networks, we feed each test CXR sample X into the trained network. Then, we obtain the network prediction mask Ŷ , which is used to generate an infection map that is a measure of infected region probabilities on the input X . Each pixel in Ŷ is defined as Ŷ h,w ∈ [0, 1] , where h and w represent the size of the image. We then apply an RGB-based color transform, i.e., the jet color scale to obtain the RGB version of the prediction mask, Ŷ R,G,B as shown in Fig. 5 for a pseudo-colored probability measure visualization. The (4)

Experimental results
In this section, first, the experimental setup is presented. Then, both numerical and visual results are reported with an extensive set of comparative evaluations over the benchmark QaTa-COV19 dataset. Finally, visual comparative evaluations are presented between the infection maps and the activation maps extracted from state-ofthe-art deep models.

Experimental setup
Quantitative evaluations for the proposed approach are performed for both COVID-19 infected region segmentation and COVID-19 detection. COVID-19 infected region segmentation is evaluated on a pixel-level, where where F-score is defined by the weighting parameter β . The F1-Score is calculated with β = 1 , which is the harmonic average of precision and sensitivity. The F2-score is calculated with β = 2 , which emphasizes FN minimization over FPs. The main objective of both COVID-19 segmentation and detection is to maximize sensitivity with a reasonable specificity in order to minimize FP COVID-19 cases or pixels. Equivalently, maximized F2-score is targeted with an acceptable F1-Score value. The performance with 95% confidence interval (CI) for COVID-19 detection is given in Table 3. The range of values can be calculated for each performance as follows: where z is the level of significance, metric is any performance evaluation metric, and N is the number of samples. Accordingly, z is set to 1.96 for 95% CI.
We have implemented the deep networks with Tensorflow library [65] using Python on NVidia ® GeForce RTX 2080 Ti GPU card. For training, Adam optimizer [66] is used with the default momentum parameters, β1 = 0.9 and β2 = 0.999 using the aforementioned hybrid loss function. The segmentation networks are trained with 50-epochs with a learning rate of α = 10 −4 and a batch size of 32.
For comparing the computed infection maps, the activation maps are computed as follows: the encoder structures of the segmentation networks are trained for the classification task with a modification at the output layer by adding 2-neurons for the number of total classes. The activation maps extracted from the classification models are then compared with the infection maps of the segmentation models. The classification networks, CheXNet, DenseNet-121, Inception-v3 and ResNet-50 are fine-tuned using categorical cross-entropy as loss function with 10 epochs and a learning rate of α = 10 −5 , which is a sufficient setting to prevent over-fitting, based on our previous study [34]. Other settings of the classifiers are kept the same with the segmentation models.

Experimental results
The experiments are carried out for both COVID-19 infected region segmentation and COVID-19 detection. We extensively tested the benchmark QaTa-COV19 dataset using three different state-of-the-art segmentation networks with four different encoder options for the initial dataset consisting of control Group-I. We also investigated the effect of frozen encoder weights on the performance. On the other hand, the leading model is selected and evaluated on the extended dataset, which includes more negative samples with the control Group-II.

Group-I experiments
We have evaluated the networks in a stratified 5-fold cross-validation scheme with a ratio of 80% training to 20% test (unseen folds) over the benchmark QaTa-COV19 dataset. The input CXR images are resized to 224 × 224 pixels. Table 1 shows the number of CXRs per fold in the dataset. Since the two classes are imbalanced, we have applied data augmentation in order to balance the classes. Therefore, COVID-19 samples are augmented up to the same number of samples as the control Group-I in the training set for each fold. The data augmentation is performed using Image Data Generator in Keras: the CXR samples are augmented by randomly shifting them both vertically and horizontally by 10% and randomly rotating them in a range of 10 degrees. After shifting and rotating the images, blank sections are filled using the nearest mode. The performance of the segmentation models for COVID-19 infected region segmentation are presented in Table 2. Each model structure is evaluated with two configurations: frozen and not frozen encoder layers. We have used transfer learning on the encoder layers with ImageNet weights, except for the CheXNet model, which is pre-trained on the ChestX-ray14 dataset. The evaluation of the models with frozen encoder layers is also important since this process can lead to a better convergence and improved performance. However, as the results show, better performance is obtained when the network continues to learn on the encoder layers as well. For each model, we have observed that two encoders: DenseNet-121 and Inception-v3 are the top-performing ones for the infected region segmentation task. The U-Net model with DenseNet-121 encoder holds the leading performance by 84% sensitivity, 85.81% F1-Score, and 84.71% F2-Score. DenseNet-121 produces better results compared to other encoder types since it can preserve the information coming from earlier layers through the output by concatenating the feature maps from each dense layer. However, in the other segmentation models, Inception-v3 outperforms the other encoder types. The presented segmentation performances are obtained by

Table 2 Average performance metrics (%) for COVID-19 infected region segmentation computed on the Group-I test (unseen) set from 5-folds with three state-of-the-art segmentation models, four encoder architectures, and weight initializations
The initialized encoder layers are set to frozen (✓) and not frozen (×) states during the investigation setting the threshold value to 0.5 to compute the segmentation mask from the network probabilities. Precision-Recall curves of three leading deep models are plotted in Fig. 6a by varying this threshold value. Additionally, the Receiver Operating Characteristics (ROC) curve of these models, and their corresponding area under curve (AUC) scores are presented in Fig. 6b. Further investigation shows that AUC scores of the leading segmentation models are directly proportional to the COVID-19 detection performance as it can be depicted from the precision scores in Table 3. The performances of the segmentation models for COVID-19 detection are presented in Table 3. All the models are evaluated by stratified a 5-fold cross-validation scheme, and the table shows the averaged results of these folds. The most crucial metric here is the sensitivity since missing any patient with COVID-19 is critical. In fact, the results indicate the robustness of the model as the proposed approach can achieve high sensitivity levels of 98.37% with a 97.08% F2-Score. Additionally, the proposed approach achieves an elegant specificity of 99.16%, indicating a significantly low false alarm rate. It can be observed from Table 3 that DenseNet-121 encoder with the not frozen encoder layer setting gives the most promising results among the others. The confusion matrices, accumulated on each fold's test set, are presented in Table 4. The highest sensitivity in COVID-19 detection is achieved by the U-Net DenseNet-121 model (Table 4a). Accordingly, the U-Net DenseNet-121 model only misses 48 COVID-19 patients out of 2951. On the other hand, the highest specificity is achieved by UNet++ DenseNet-121 model (Table 4b). The UNet++ model only misses a minor part of the control class with 105 samples out of 12544.

Group-II experiments
We have selected the leading model from the Group-I experiments as U-Net with not frozen DenseNet-121 encoder setting. In Group-II experiments, we have gathered around 120K CXRs. The CXRs from the ChestX-ray14 dataset [55] are already divided into train and test sets. Accordingly, we have randomly separated the train and test sets of COVID-19, viral pneumonia, and bacterial pneumonia CXRs by keeping the same train/test ratio as in ChestX-ray14 [55].

Infection vs activation maps
Several studies [36][37][38] propose to localize COVID-19 from CXRs by extracting activation maps from the deep classification models trained for COVID-19 detection.    Table 7 Cumulative confusion matrices of COVID-19 detection by the best performing U-Net model with DenseNet-121 encoder

Group-II COVID-19
Ground Truth Group-II 26, 534 31 COVID- 19 44 Despite the simplicity of the idea, there are many limitations of this approach. First of all, without any infected region segmentation ground-truth masks, the network can only produce a rough localization, and the extracted activation maps may entirely fail to localize COVID-19 infection.
In this study, we check the reliability of our proposed COVID-19 detection approach by comparing it with DL models trained for the classification task. In order to achieve this objective, we compare the infection map and activation map of CXR images, which are generated from the segmentation and classification networks, respectively. Therefore, we have trained the encoder structures of the segmentation networks, which are CheXNet, DenseNet-121, Inception-v3, and ResNet-50 to perform COVID-19 classification task. We have extracted activation maps from these trained models by the Gradient-weighted Class Activation Mapping (Grad-CAM) approach proposed in [67]. The localization Grad-CAM L c Grad-CAM ∈ R h×w of height h and width w for class c is calculated by the gradient of m c before the softmax with respect to the convolutional layer's feature maps A k as ∂m c ∂A k . The gradients are passed through from the global average pooling during back-propagation; where α is the weight that shows the important feature map k from A for a target class c. Then, the linear combination is performed following by ReLU to obtain the Grad-CAM; Despite their elegant performance, activation maps extracted from deep classification networks are not suitable for localizing COVID-19 infection as depicted in Fig. 8  implementation of the infection maps will obviously speed up the detection process, can also monitor the progression of COVID-19 infection in the lungs.

Computational complexity analysis
In this section, we present the computational times of the networks and their number of trainable & non-trainable parameters. Table 8 shows the elapsed time in milliseconds (ms) during the inference step for each network used in the experiments. The results in the table represent the running time per sample. It can be observed from the table that the U-Net model is the fastest among the others due to its shallow structure. The fastest network is U-Net Inception-v3 with frozen encoder layers taking up 2.53 ms. On the other hand, the slowest model is UNet++ structure since it has the largest number of trainable parameters. The most computationally demanding model is UNet++ ResNet-50 with frozen encoder layers, which takes 5.58 ms. We, therefore, conclude that all models can be used as real-time clinical applications.

Conclusions
The immediate and accurate detection of highly infectious COVID-19 plays a vital role in preventing the spread of the virus. In this study, we used CXR images since X-ray imaging is cheaper, easily accessible, and faster than the conventional methods commonly used such as RT-PCR and CT. As a major contribution, the largest CXR dataset, QaTa-COV19, which consists of 2951 COVID-19, and 116, 365 control group images, has been compiled and is shared publicly as a benchmark dataset. Moreover, for the first time in the literature, we release the ground-truth segmentation masks of the infected regions along with the introduced benchmark QaTa-COV19. Furthermore, we proposed a humanmachine collaborative approach, which can be used when a fast and accurate ground-truth annotation is desired but manual segmentation is slow, costly, and subjective. Finally, this study reveals the first approach ever proposed for infection map generation in CXR images. Our extensive experiments on QaTa-COV19 show that a reliable COVID-19 diagnosis can be achieved by generating infection maps, which can locate the infection on the lungs by 81.72% sensitivity, and 83.20% F1-Score. Moreover, the proposed joint approach can achieve an elegant COVID-19 detection performance with 94.96% sensitivity and 99.88% specificity. Many COVID-19 detectors proposed in the literature reported similar or even better detection performances. However, not only they are evaluated over small-size datasets, but also they can only discriminate between COVID-19 and normal (healthy) data, which is a straightforward task. The proposed joint approach is the only COVID-19 detector that can distinguish it from other thoracic diseases as being evaluated over the largest CXR dataset ever composed. Accordingly, the most important aspect of this study is that the generated infection maps can assist MDs for a better and objective COVID-19 assessment. For instance, it can show the time progress of the disease if the time series CXR data are generated by the proposed infection maps. It is clear that when compared with the activation maps extracted from deep models, the proposed infection maps are highly superior and reliable cues for COVID-19 infection.