key: cord-0139399-51o7s97r authors: Hasan, Md. Kamrul; Jawad, Md. Tasnim; Hasan, Kazi Nasim Imtiaz; Partha, Sajal Basak; Masba, Md. Masum Al title: COVID-19 identification from volumetric chest CT scans using a progressively resized 3D-CNN incorporating segmentation, augmentation, and class-rebalancing date: 2021-02-11 journal: nan DOI: nan sha: 4fa7f445dce2b24c8827722bb0082194f4e91e8e doc_id: 139399 cord_uid: 51o7s97r The novel COVID-19 is a global pandemic disease overgrowing worldwide. Computer-aided screening tools with greater sensitivity is imperative for disease diagnosis and prognosis as early as possible. It also can be a helpful tool in triage for testing and clinical supervision of COVID-19 patients. However, designing such an automated tool from non-invasive radiographic images is challenging as many manually annotated datasets are not publicly available yet, which is the essential core requirement of supervised learning schemes. This article proposes a 3D Convolutional Neural Network (CNN)-based classification approach considering both the inter- and intra-slice spatial voxel information. The proposed system is trained in an end-to-end manner on the 3D patches from the whole volumetric CT images to enlarge the number of training samples, performing the ablation studies on patch size determination. We integrate progressive resizing, segmentation, augmentations, and class-rebalancing to our 3D network. The segmentation is a critical prerequisite step for COVID-19 diagnosis enabling the classifier to learn prominent lung features while excluding the outer lung regions of the CT scans. We evaluate all the extensive experiments on a publicly available dataset, named MosMed, having binary- and multi-class chest CT image partitions. Our experimental results are very encouraging, yielding areas under the ROC curve of 0.914 and 0.893 for the binary- and multi-class tasks, respectively, applying 5-fold cross-validations. Our method's promising results delegate it as a favorable aiding tool for clinical practitioners and radiologists to assess COVID-19. exponentially. The incubation period, which is a time between catching the virus and causing to have indications of the illness, is 1 ∼ 14 days, making it remarkably challenging to identify COVID-19 infection at a preliminary stage of an individual's symptoms [48] . The clinical screening test for the COVID-19 is Reverse Transcription Polymerase Chain Reaction (RT-PCR), practicing respiratory exemplars. However, it is a manual, complicated, tiresome, and time-consuming fashion with an estimated true-positive rate of 63.0% [103] . There is also a significant lack of RT-PCR kit inventory, leading to a delay in preventing and curing coronavirus disease [112] . Furthermore, the RT-PCR kit is estimated to cost around 120 ∼ 130 USD. It also requires a specially designed biosafety laboratory to house the PCR unit, each of which can cost 15, 000 ∼ 90, 000 USD [1] . Nevertheless, the utilization of a costly screening device with a delayed test results makes it more challenging to restrict the disease's spread. Inadequate availability of screening workstations and measurement kits constitute an enormous hardship to identify COVID-19 in this pandemic circumstance. In such a situation, speedy and trustworthy presumed COVID-19 cases are an enormous difficulty for related personals. However, it is observed that most of the COVID-19 incidents have typical properties on radiographic CT and X-ray images, including bilateral, multi-focal, ground-glass opacities with a peripheral or posterior distribution, chiefly in the lower lobes and early-and late-stage pulmonary concentration [18, 42, 88, 110] . Those features can be utilized to build a sensitive Computer-aided Diagnosis (CAD) tool to identify COVID-19 pneumonia, which is deemed an automated screening tool [59] . Currently, deep Convolutional Neural Networks (CNNs) allow for building an end-to-end model without requiring manual and time-consuming feature extraction and engineering [57, 58] , demonstrating tremendous success in many domains of medical imaging, such as arrhythmia detection [4, 28, 113] , skin lesion segmentation and classification [17, 23, 24, 35] , breast cancer detection [13, 19, 31] , brain disease segmentation and classification [93, 97] , pneumonia detection from chest X-ray images [79] , fundus image segmentation [34, 94] , and lung segmentation [26] . Most recently, various deep CNN-based methods have been published for identifying COVID-19 from X-rays and CT images, summarizing and bestowing in Table 1 . Though the results obtained in the current articles are Table 1 : Numerous published articles for the COVID-19 identification with their respective utilized datasets and performances exhibiting different metrics such as mSn, mSp, and mF1 respectively for mean sensitivity, specificity, and F1-score. The mixed datasets indicate that data have come from different open-sources. A pre-trained 2D MobileNet-v2 [82] architecture on ImageNet [20] was used to extract massive high-dimensional features to classify six different diseases using the fully-connected layers [7] Mixed mSn: 0.974 mSp: 0.994 DeTraC [2, 3] , where the network was trained first using a gradient descent optimization [81] , and then, the class-composition layer of DeTraC was used to refine the final detection results [2] Mixed mSn: 0.979 mSp: 0.919 A multi-objective differential evolution-based CNN method fine-tuning iteratively using mutation, crossover, and selection operations to discover the best possible results [88] Mixed mSn: 0.907 mSp: 0.906 An ensemble of VGG-16 [87] , Inception [92] , Xception [16] , Inception-ResNet [91] , MobileNet [41] , DenseNet [43] , and NasNet [77] optimizing the hyperparameters using a greedy search algorithm [10, 78] promising, they exhibit limited scope for use as a CAD tool, as most of the works, especially on x-ray images, have been based on data coming from different sources for two distinct classes (Covid Vs. Normal) [8, 38, 53, 67, 74, 84, 88, 102] . This brings inherent bias on the algorithms as the model tends to learn the distribution of the data source for binary classification problem [32] . Therefore, these models perform very low when used in practical settings, where the models have to adapt to data from different domains [32] . • Designing a 3D-CNN-based classification network for volumetric CT images as the 3D networks account for the inter-and intra-slice spatial voxel information while the 2D networks consider only the intra-slice spatial voxel information [37, 44, 52, 89, 114, 118] • Conducting 3D patch-based classification as it increases the sample numbers in the smaller datasets, where we perform ablation studies to determine a proper patch size • Progressively increasing the input patch size of our network up to the original CT size of R × C × S, where the trained network with the patch size of (R/2 n+1 ) × (C/2 n+1 ) × (S/2 n+1 ) is a pre-trained model of a network with the patch size of (R/2 n ) × (C/2 n ) × (S/2 n ) • Developing an unsupervised lung segmentation pipeline for allowing the classifier to learn salient lung features while omitting the outer lung areas of the CT scans • Class rebalancing and augmentations, such as intensity-and geometry-based, are employed to develop a general network, although a small dataset is being utilized The remainder of the article is prepared as follows. Section 2 details the materials and methods practiced in the study, including a brief introduction to the methodology and end-to-end 3D-CNN training. Section 3 describes the experimental operations and their corresponding obtained results. Lastly, section 4 concludes the article. In this section, we describe the utilized materials and methods to conduct the widespread experiments. We summarize the adopted dataset in the first subsection 2. Table 2 illustrates that the class distribution is imbalanced. Such an imbalanced class distribution produces a biased image classifier towards the class having more training samples. We apply various rebalancing schemes to develop a generic classifier for COVID-19 identification, even though the dataset is imbalanced. The recommended integral preprocessing consists of segmentation, augmentations (both geometry-and intensity-based), and class-rebalancing, which are concisely explained as follows: Segmentation. The segmentation, to separate an image into regions with similar properties such as gray level, color, texture, brightness, and contrast, is the significant element for automated detection pipeline [39] . It is also a fundamental prerequisite for the COVID- primary step is transforming all the CT volumes to Hounsfield units (HU), as it is a quantitative measure of radiodensity for CT scans. We set the HU unit as -1000 to -400 as the study shows that lung regions are within that range, which was also used in many articles [56, 85, 101] . The thresholded binary lung masks are then refined to exclude dif- Rebalancing. The utilized dataset in Table 2 is imbalanced. This situation is pretty obvious in the medical diagnosis field due to the scarcity of massive manually annotated training samples, especially in COVID-19 datasets. The undesired class-biasing occurs in the supervised learning systems towards the class with majority samples. However, we apply two techniques to rebalance the imbalanced class distribution, such as adding extra CT volumes from the publicly available CC-CCII dataset [115] and weighting the loss function for penalizing the overrepresented class. The latter approach rewards more extra consideration to the class with minority samples. Here, we estimate the class weight using a portion of W n = N n /N , where W n , N , and N n separately denote the n th -class weight, the total sample numbers, and the samples in n th -class. We employ both the class-rebalancing strategies in the binary-class protocol, whereas the only class weighting method is adopted in the multi-class protocol. The deep neural network is a machine learning framework with a wide range of applications, from natural language processing [21] to medical image classification [12] , segmentation [12] , and registration [25] . In special, CNNs have become a prevalent technique in the computer vision community. They are practiced in diverse tasks, including object detection [50] , classification [22] , and localization [63] . The CNN-based deep neural systems are also popularly adopted in recent pandemic for COVID-19 identification [11, 73] (see in Table 1 ). CNN is an excellent discriminant feature extractor at various abstraction levels, which is translation-invariant. Consequently, utilizing it to classify medical images evades complicated and expensive feature engineering [83] . The early few CNN layers learn low-level image features and later layers learn high-level image features particular to the application types [52] . However, the 2D-CNNs are frequently employed in natural RGB and grayscale images to extract the spatial features only in two dimensions [92] . The 2D-CNN also can be applied to the volumetric medical image datasets taking cross-sectional 2D slices of the CT, MRI, or similar scans. However, the recent experimental results have revealed the advantages of 3D-CNN over 2D-CNN, where the 3D-CNN accepts the volumetric spatial information as an input [62] . Conventional 2D-CNNs' effectiveness is degraded due to loss of spatial voxel information for volumetric 3D medical imaging tasks. A 3D-CNN, a 3D space implementation of convolution and pooling operation, is practiced to overcome spatial voxel information loss as in the 2D-CNNs. The image becomes scalable in the spatial direction using a 3D-CNN, allowing accurate image detection with different frame sizes [64] . Therefore, we propose a classifier based on 3D-CNN to identify COVID-19 from the volumetric CT scans. The proposed base network in Fig. 2 essentially consists of two modules, such as feature extractor and feature classifier. The former module is a stack of convolutional, pooling, and batch normalization layers, whereas the latter module is a stack of fully-connected layers followed by a softmax layer. We involve 3D layers for all the feature extractor module components to operate on volumetric medical images for extracting the most discriminating features, accounting for both the intra-and inter-slice spatial voxel information. In our network, each 3D convolutional layer with ReLU activation is followed by a 3D max-pooling layer, where the pooling layer increases translational invariances of the network. The pooled feature maps are then used as an input to the successive layers, which may dynamically change during training at each training epoch [89] . The more enormous changes prone to bring difficulties for searching an optimal parameter or hyperparameter; often become computationally expensive to reach an optimal value [47] . Such a problem is mitigated by integrating batch normalization layers in our network [47] . It also facilitates the smooth training of the network architectures in less time [89] . The Global Average Pooling (GAP) [61] is used as a bridge layer between the feature extractor and feature classifier modules, converting the feature tensor into a single long continuous linear vector. In GAP, only one feature map is produced for each corresponding category, achieving a more extreme dimensionality compression to evade overfitting [61] . A dropout layer [90] is also employed as a regulariser, which randomly sets half of the activation of the fully-connected layers to zero through the training of our network. Again, as mentioned earlier, the CNNs are heavily reliant on the massive dataset to bypass overfitting and build a generic network. The acquisition of annotated medical images is arduous to accumulate, as the medical data collection and labeling are confronted with data privacy, requiring time-consuming expert explanations [111] . There are two general resolving directions: accumulating more data, such as crowdsourcing [51] or digging into the present clinical reports [105] . Another technique is investigating how to enhance the achievement of the CNNs with small datasets, which is exceptionally significant because the understanding achieved from the research can migrate the data insufficiency in the medical imaging fields [111] . Transfer learning is a widely adopted method for advancing the performance of CNNs with inadequate datasets [15] . CT scans' size, as pictured in Fig. 3 . Such training is called progressive resizing [9] , where the training begins with smaller image sizes followed by a progressive expansion of size. This training process is continued until the last patch and network sizes are as same as the initial image dimension. We first extract five different patches with different sizes (see in Fig. 4) to begin the experimentations. We perform ablation studies in subsection 3.1 looking for the best patch size. The weights of the base network in Fig. 2 is initialized with Xavier normal distribution. The weights of the first progressively resized network are initialized with the weights of the base network. In general, the weights of the network with the patch size of (R/2 n ) × (C/2 n ) × (S/2 n ) are initialized with the weights of the network with the patch size of (R/2 n+1 ) × (C/2 n+1 ) × (S/2 n+1 ) for the original CT volume size of R × C × S. Categorical cross-entropy and accuracy are utilized as a loss function and metric, respectively, for training all the networks in this article. We use Adam [54] optimizer with initial learning rate (LR), exponential decay rates (β 1 , β 2 ) as LR = 0.0001, β 1 = 0.9, and β 2 = 0.999, respectively, without AMSGrad variant. The exponential decaying LR schedule is also employed for the networks' optimization. Initial epochs are set as 200, and training is terminated if validation performance stops growing after 15 epochs. We We evaluate all the experimental outcomes by employing numerous metrics, such as recall, precision, and F1-score, for evaluating them from diverse perspectives. The recall measures the type-II error (the patient having positive COVID-19 characteristics, erroneously abandons to be repealed), whereas the precision estimates the positive predictive values (a portion of absolutely positive-identification amid all the positive-identification). The harmonic mean of recall and precision is manifested using the F1-score, conferring the tradeoff between these two metrics. Furthermore, we also quantify the prognostication probability of an anonymously selected CT sample using a Receiver Operating Characteristics (ROC) with its Area Under the ROC Curve (AUC) value. In We extract five different 3D patches, named P 1 , P 2 , P 3 , P 4 , and CT scans having size of 512 × 512 × 36 is named as P 6 . The height and width of the patch P 5 is half of the P 6 , whereas these dimensions of the patch P 4 is one-fourth of the P 6 , and so on. We extract 2 n number of patches for a n th -time reduction of the height and width. Therefore, we train and test our network with 71040, 35520, 17760, 8880, 4440, and 1110 samples for the 3D volumes P 1 to P 6 , respectively. The examples of the extracted patches are shown in Fig. 4 , where we select the middle slices of the extracted patches of the same CT scan. Different patches in Fig. 4 shows their respective resolutions, where it is seen that the patches P 1 and P 2 demonstrate very low resolutions. However, the effects of those patch resolutions are judged by classifying the NOR vs. NCP classes (see in subsection 2.1). (a) The classification results are presented in Fig. 5 for all the patches (P 1 to P 5 ) and original CT scans (P 6 ) employing our 3D network without any type of preprocessing. The results show that the network inputting with P 1 patch outputs COVID-19 identification with type-II errors as 69.0 % and 25.0 % for NOR-and NCP-classes, respectively. Such results confirm that NCP-class has been identified more accurately (44.0 % more in NCP-class), pointing that classifier is biased towards the NCP-class. On the other hand, the utilization of patch P 2 produces identification results with type-II errors as 56.0 % and 39.0 % for NOR-and NCPclasses, which reduce the differences between two classes (only 17.0 % more in NCP-class). Although the P 1 patch has double samples, it fails to provide a class-balanced performance than the P 2 patch. This is because of having a better-resolution in the P 2 patch than the P 1 patch (see in Fig. 4) , as other experimental settings are constant. Again, the patch P 3 further improves the identification results with type-II errors as 54.0 % and 28.0 % for NORand NCP-classes. Approximately, the patch P 4 also provides similar results to the P 3 patch. It is noteworthy from those experimentations that P 3 or P 4 patches have much fewer samples than P 1 (4-times and 8-times, respectively); still, they outperform the identification results of P 1 and P 2 patches with the same experimental settings. Furthermore, the utilization of patch P 5 further reduces the performances (type-II errors as 6.0 % and 99.0 % for NOR-and NCP-classes) than all the previous patches discussed above. Such a result shows that it produces a more biased model towards the NCP-class. From Fig. 4 shows that the patch P 4 and P 5 are visually looking similar but P 4 has twotimes samples as of P 5 . This experiment exposes that having fewer samples also generates class-biased classifiers if input images are similar in resolution. Finally, the network with the original images also provides less COVID-19 identification performance as in the patch P 5 (see in Fig. 5 ). All the experiments show that our network with P 3 or P 4 patches has outputted better-identification results. Such experimental results undoubtedly prove that both the input resolution and the number of samples play an important role in CNN-based classifiers. We can not increase the number of samples taking the smaller patch sizes, as it has a shallow resolution, which adversely affects the classifiers. The aforementioned results reveal that the utilization of better-resolution with more sample numbers increases the performance of CNN. Therefore, we propose to employ progressive resizing of our proposed 3D-CNN (see details in subsection 2.3). Firstly, we begin training our network with a suitable 3D patch with more training samples from the previous experiments, acting as a base model. Then, we add some CNN layers to the input of the base model with the higher resolution (2-times more in this article), where the base model is adopted as a pre-trained model (see details in subsection 2.3). We repeat this network resizing until we reach to original given CT size (P 6 ). The results for such a progressive resizing are presented in the confusion metrics in Table 3 and ROC curves (with respective AUC values) in Fig. 6 . The confusion matrix in Table 3 , Table 3 : Normalized confusion matrix employing our network with progressive resizing, where we progressively increase the input resolution from P 4 to P 5 then to P 6 (original resolution). The first table (left) for the resolution of P 4 , the second table (middle) for resolution of P 4 −→ P 5 , and the last (right) for resolution of P 4 −→ P 5 −→ P 6 . resizing (P 4 −→ P 5 ). Again, employing P 4 −→ P 5 as a pre-trained model, the utilization of P 6 (original CT scans), with 1110 samples, increases the false-negative rate of NCP by 3.92 %, still less than baseline false-negative rate of 13.52 %. It also decreases the falsepositive rate by a margin of 18.14 %, which is less than the former two false-positive rates (see all tables in Table 3 ). Furthermore, the proposed final progressively resized network (P 4 −→ P 5 −→ P 6 ) obtains an AUC of 0.754, which indicates that the probability of correct COVID-19 identification is as high as 75.4 % for any given random CT samples (see in Fig. 6 ). It has beaten the baseline P 4 and P 4 −→ P 5 respectively by 17.0 % and 7.70 % in terms of AUC, as presented in Fig. 6 . Although the final progressively resized network (P 4 −→ P 5 −→ P 6 ) has an input of the original CT scans, its performance is far better than the network training with P 6 alone (see in Fig. 5 ). All the above discussions in this subsection experimentally validate the progressive resizing supremacy for the COVID-19 identification instead of training using single size input CT scans. This subsection presents the COVID-19 identification results from our progressively resized 3D network employing different preprocessing, such as augmentation, segmentation, and class-rebalancing. Segmentation. The well-defined segmentation, with less-coarseness, is an essential requirement for further identification. The incorporation of segmentation with the PRN further promotes the identification results than the PRN alone, as exposed in Table 4 . Several examples of the segmented lung from our proposed unsupervised pipeline (as described in subsection 2.2) are depicted in Fig. 7 for qualitative evaluation. However, the COVID-19 identification results incorporating lung segmentation with the PRN reflects its supremacy over the PRN alone, extending the weighted average type-II error by 1.4 % with respective improvements in average positive predictive value by 2.8 % (see in Table 4 ). The classimbalanced identification is also dwindled due to segmented lung area utilization over the full CT volumes. The reasonable ground for those enhanced performances due to the segmentation is that it extracts an abstract region, enabling the classifier to learn only the precise lung areas' features while avoiding the surrounding healthy tissues of the chest CT scans. Augmentation, Segmentation, and Class-rebalancing. The combination of augmentations, segmentation, and class-rebalancing with the PRN provides the best COVID-19 identification results of this article. This experiment identifies the COVID-19 from the chest CT scans with relatively less class-imbalance with the weighted average type-II error of 13.0 % with respective average positive predictive value as 13.1 %. All the preprocessing employment heightens the former metric by a margins of 8.1 % and the latter metric by 9.7 % from the baseline model (see in Table 4 ) with less class-imbalance performance. Besides, Fig. 8 This subsection displays the COVID-19 identification results using our proposed PRNASCR for binary-and multi-class (see in subsection 2.1) utilizing the 5-fold cross-validation. The detailed class-wise performance of our PRNASCR for both the binary-and multi-class is exhibited in the confusion metrics in Table 5 (left) and Table 5 (right), correspondingly. The binary-classification results in Table 5 ). Although overall macro-average AUC of the binary classification defeats the multi-class recognition (see in multi-class protocol also provides less inter-fold variation than the binary-class, as depicted in Fig. 9 . However, our approach for the COVID-19 identification exhibits praiseworthy achievement with high AUC values with less inter-fold variation in both of the class protocols. During the current COVID-19 pandemic emergency, to mitigate the permanent lung damage due to coronavirus, precise recognition with negligible false negative is highly essential. This article aimed to design an artificial screening system for automated COVID-19 identification. A progressively resized 3D-CNN classifier is recommended in this study, incorporating lung segmentation, image augmentations, and class-rebalancing. The experimental analysis confirms that the CNN classifier's training with the suitable smaller patches and progressively increasing the network size enhance the identification results. Furthermore, incorporating the lung segmentation empowers the classifier to learn salient and characteristic COVID-19 features than utilizing whole chest CT images, driving to improved COVID-19 classification performance. Again, the augmentations and class-rebalancing result in im-proved COVID-19 identification with high class-balanced recognition, shielding the network from being biased to a particular overrepresented class. In the future, the proposed pipeline will be employed in other volumetric medical imaging domain to validate its efficacy, versatility, and robustness. We also aim to deploy our trained model to a user-friendly web application for clinical utilization. The proposed system can be an excellent tool for clinicians to fight this deadly epidemic by the quicker and automated screening of the COVID-19. Bangladesh scientists create $3 kit. Can it help detect COVID-19 Classification of covid-19 in chest x-ray images using detrac deep convolutional neural network Detrac: Transfer learning of class decomposed medical images in convolutional neural networks A deep convolutional neural network model to classify heartbeats Explainable covid-19 detection using chest ct scans and deep learning Explainable-by-design approach for covid-19 classification via ct-scan. medRxiv Extracting possibly representative covid-19 biomarkers from x-ray images with deep learning approach and image data related to pulmonary diseases Covid-19: automatic detection from x-ray images utilizing transfer learning with convolutional neural networks Rgpnet: A real-time general purpose semantic segmentation Random search for hyper-parameter optimization Deep learning and medical image processing for coronavirus (covid-19) pandemic: A survey. Sustainable cities and society 65 A review of the application of deep learning in medical image classification and segmentation. Annals of translational medicine 8 Automated invasive ductal carcinoma detection based using deep transfer learning with whole-slide images Epidemiological and clinical characteristics of 99 cases of 2019 novel coronavirus pneumonia in wuhan, china: a descriptive study Not-so-supervised: a survey of semi-supervised, multi-instance, and transfer learning in medical image analysis Xception: Deep learning with depthwise separable convolutions Deep learning ensembles for melanoma recognition in dermoscopy images Detection of 2019 novel coronavirus (2019-ncov) by real-time rt-pcr Automatic detection of invasive ductal carcinoma in whole slide images with convolutional neural networks, in: Medical Imaging Imagenet: A large-scale hierarchical image database Deep learning in natural language processing Image classification using convolutional neural network (cnn) and recurrent neural network (rnn): A review Skin lesion classification using convolutional neural network for melanoma recognition. medRxiv Dermatologist-level classification of skin cancer with deep neural networks Deep learning in medical image registration: a review Attention u-net based adversarial architectures for chest x-ray lung segmentation Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow: Concepts, tools, and techniques to build intelligent systems Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network Prediction of epileptic seizure by analysing time series eeg signal using-nn classifier Cvr-net: A deep convolutional neural network for coronavirus recognition from chest radiography images Automatic mass detection in breast using deep convolutional neural network and svm classifier Challenges of deep learning methods for covid-19 detection using public datasets Diabetes prediction using ensembling of different machine learning classifiers Drnet: Segmentation and localization of optic disc and fovea from diabetic retinopathy image DSNet: Automatic dermoscopic skin lesion segmentation Deep residual learning for image recognition Benchmarking deep learning models and automated model design for covid-19 detection with chest ct scans Covidx-net: A framework of deep learning classifiers to diagnose covid-19 in x-ray images Deep learning techniques for medical image segmentation: Achievements and challenges Covid-19 detection through transfer learning using multimodal imaging data Mobilenets: Efficient convolutional neural networks for mobile vision applications Clinical features of patients infected with 2019 novel coronavirus in wuhan, china Densely connected convolutional networks Lung nodule detection in ct using 3d convolutional neural networks Differential data augmentation techniques for medical imaging classification tasks Squeezenet: Alexnet-level accuracy with 50x fewer parameters and¡ 0.5 mb model size Batch normalization: Accelerating deep network training by reducing internal covariate shift A deep learning approach to detect covid-19 coronavirus with x-ray images Classification of the covid-19 infected patients using densenet201 based deep transfer learning Cnn-based encoder-decoder networks for salient object detection: A comprehensive review and recent advances Capsule networks against medical imaging data challenges, in: Intravascular Imaging and Computer Assisted Stenting and Large-Scale Annotation of Biomedical Data and Expert Label Synthesis Deep 3d convolution neural network for ct brain hemorrhage classification Coronet: A deep neural network for detection and diagnosis of covid-19 from chest x-ray images Adam: A method for stochastic optimization Covid-19 pneumonia diagnosis using a simple 2d deep learning framework with a single chest ct image: model development and validation Chest ct: automated nodule detection and assessment of change over time-preliminary experience Imagenet classification with deep convolutional neural networks Deep learning Covid-19 pneumonia: what has ct taught us? Early transmission dynamics in wuhan, china, of novel coronavirus-infected pneumonia A survey on deep learning in medical image analysis Accurate object localization in remote sensing images based on convolutional neural networks A 3d convolutional neural network for volumetric image semantic segmentation Covtanet: A hybrid tri-level attention based network for lesion segmentation, diagnosis, and severity prediction of covid-19 chest ct scans Mosmeddata: Chest ct scans with covid-19 related findings Automatic detection of coronavirus disease (covid-19) using x-ray images and deep convolutional neural networks Application of deep learning techniques for detection of covid-19 cases using chest x-ray images: A comprehensive study A novel medical diagnosis model for covid-19 infection detection based on deep features and bayesian optimization Ensemble-cvdnet: A deep learning based end-to-end classification framework for covid-19 detection using ensembles of networks Cvdnet: A novel deep learning architecture for detection of coronavirus (covid-19) from chest x-ray images Coronavirus (covid-19) classification using deep features fusion and ranking technique Review on diagnosis of covid-19 from chest ct images using artificial intelligence Automated detection of covid-19 cases using deep neural networks with x-ray images Deep transfer learning based classification model for covid-19 disease Covid-19 identification in chest x-ray images on flat and hierarchical classification scenarios Efficient neural architecture search via parameter sharing Iteratively pruned deep learning ensembles for covid-19 detection in chest x-rays Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning U-net: Convolutional networks for biomedical image segmentation An overview of gradient descent optimization algorithms Proceedings of the IEEE conference on computer vision and pattern recognition Convolutional neural networks in medical image understanding: a survey Detection of coronavirus disease (covid-19) based on deep features Automatic lung segmentation in ct images using watershed transform A survey on image data augmentation for deep learning Very deep convolutional networks for large-scale image recognition Classification of covid-19 patients from chest ct images using multi-objective differential evolution-based convolutional neural networks 2020b. 3d deep learning on medical images: a review Dropout: a simple way to prevent neural networks from overfitting Inception-v4, inception-resnet and the impact of residual connections on learning Going deeper with convolutions Convolutional neural networks for multi-class brain disease detection using mri images Automated segmentation of exudates, haemorrhages, microaneurysms using single convolutional neural network Efficientnet: Rethinking model scaling for convolutional neural networks Covid-19 detection using deep learning models to exploit social mimic optimization and structured chest x-ray images using fuzzy color and stacking approaches Brain tissue segmentation using neuronet with different pre-processing techniques Covidgan: Data augmentation using auxiliary classifier gan for improved covid-19 detection The impact of covid-19 and strategies for mitigation and suppression in low-and middle-income countries Clinical characteristics of 138 hospitalized patients with 2019 novel coronavirus-infected pneumonia in wuhan, china Automated segmentation of lungs with severe interstitial lung disease in ct Covid-19 classification by fgcnet with deep feature fusion from graph convolutional network and convolutional neural network Detection of sars-cov-2 in different types of clinical specimens A weaklysupervised framework for covid-19 classification and lesion localization from chest ct Tienet: Text-image embedding network for common thorax disease classification and reporting in chest x-rays Combining support vector machine with genetic algorithm to classify ultrasound breast tumor images The outbreak of covid-19: An overview Empirical evaluation of rectified activations in convolutional network Deep convolutional neural network based medical image classification for disease diagnosis Point-of-care rna-based diagnostic device for covid-19 Arrhythmia detection using deep convolutional neural network with long duration ecg signals Performance and robustness of machine learning-based radiomic covid-19 severity prediction Clinically applicable ai system for accurate diagnosis, quantitative measurements, and prognosis of covid-19 pneumonia using computed tomography Shufflenet: An extremely efficient convolutional neural network for mobile devices Covid-ct-dataset: a ct scan dataset about covid-19 Performance evaluation of 2d and 3d deep learning approaches for automatic segmentation of multiple organs on ct images Learning transferable architectures for scalable image recognition None. No funding to declare. All authors have no conflict of interest to publish this research.