key: cord-0699188-3ov501xx authors: Sufian, Abu; Ghosh, Anirudha; Sadiq, Ali Safaa; Smarandache, Florentin title: A Survey on Deep Transfer Learning and Edge Computing for Mitigating the COVID-19 Pandemic date: 2020-06-30 journal: nan DOI: 10.1016/j.sysarc.2020.101830 sha: 29f492297b65015d2d24af5db104f46faa315263 doc_id: 699188 cord_uid: 3ov501xx Abstract Global Health sometimes faces pandemics as are currently facing COVID-19 disease. The spreading and infection factors of this disease are very high. A huge number of people from most of the countries are infected within four months from its first report of appearance and it keeps spreading. The required systems are not ready up to some stages for any pandemic; therefore, mitigation with existing capacity becomes necessary. On the other hand, modern-era largely depends on Artificial Intelligence(AI) including Data Science, and Deep Learning(DL) is one of the current flag-bearer of these techniques. It could use to mitigate COVID-19 like pandemics in terms of stop spread, diagnosis of the disease, drug & vaccine discovery, treatment, and many more. But this DL requires large datasets as well as powerful computing resources. A shortage of reliable datasets of a running pandemic is a common phenomenon. So, Deep Transfer Learning(DTL) would be effective as it learns from one task and could work on another task. In addition, Edge Devices(ED) such as IoT, Webcam, Drone, Intelligent Medical Equipment, Robot, etc. are very useful in a pandemic situation. These types of equipment make the infrastructures sophisticated and automated which helps to cope with an outbreak. But these are equipped with low computing resources, so, applying DL is also a bit challenging; therefore, DTL also would be effective there. This article scholarly studies the potentiality and challenges of these issues. It has described relevant technical backgrounds and reviews of the related recent state-of-the-art. This article also draws a pipeline of DTL over Edge Computing as a future scope to assist the mitigation of any pandemic. The COVID-19 is a disease caused by a novel coronavirus called 'SARS-CoV-2'. This virus is transferable from human to human and its spreading, and infection factors are very high [1, 2] . Almost three million people are infected and over 200 thousands are died within just four months from 5 it's originating, and it is increasing steadily 1 . The World Health Organization(WHO) has declared it a pandemic [3, 4] . But this is not the only pandemic human civilization is facing, there are many outbreaks had come in the past or it may come in the future [5, 6] . The appropriate drugs, vaccines, infrastructure, etc. are not available up to some stages of any out- 10 breaks. Therefore, mitigate these types of diseases with existing capacity becomes most important in those stages [7, 8] . Many researchers from all over the world trying hard to develop such kind of techniques to cope with such challenges [9, 10] . Modern-era largely depends on Artificial Intelligence(AI) including Data 15 Science, and Deep Learning(DL) is one of the current flag-bearer of these techniques [11] . Therefore, these techniques could also assist to mitigate COVID-19 like pandemics in terms of stop spread, diagnosis of the disease, drug & vaccine discovery, treatment, and many more [12, 13] . But to trained this DL, large datasets as well as powerful computing resources are required. 20 For a new pandemic, data insufficiency and it's variation over different geographic regions is a huge problem, so here Deep Transfer Learning (DTL) would be effective as it learns from one task and could apply in another task after required fine-tuning [14] . On the other hand edge devices such as IoT, Webcam, Drone, Intelligent Medical Equipment, Robot, etc. are 25 very useful in any pandemic situation. These types of equipment make the infrastructures sophisticated and automated which helps to cope with an outbreaks [15] . Though, such devices are equipped with low computing resources which represent the main challenges of Edge Computing(EC) [16] . As a way to overcome this challenge, transfer learning could be a possible way 30 to consolidate the needed computational power and facilitate more efficient EC. Therefore, DTL in edge devices as an EC could be smart techniques to mitigate a new pandemic [17] . This survey article has tried to report all these issues scholarly as potentialities and challenges with relevant technical backgrounds. Here, we also proposed a possible pipeline architecture for 35 future scopes to brings DTL over EC to assists mitigation in any outbreaks. 1 https://www.worldometers.info/coronavirus/ Some highlights of the contributions of this article are as follows: • Presented a systematic study of Deep Learning(DL), Deep Transfer Learning(DTL) and Edge Computing(EC) to mitigate COVID-19. 40 • Surveyed on existing DL, DTL, EC, and Dataset to mitigate pandemics with potentialities and challenges. • Drawn a precedent pipeline model of DTL over EC for a future scope to mitigate any outbreaks. • Given brief analyses and challenges wherever relevant in perspective of 45 COVID-19. Starting from the introduction in section 1, the remainder of the article organized follows. Section 2 for technical background whereas review of generic state-of-the-art of DTL in EC in section 3. Existing computing(DL, 50 DTL, EC & Dataset) to mitigate pandemic in section 4. A proposed pipeline of DTL in EC to mitigate pandemics in section 5. Finally, conclusion in section 6. The main focus of this article is how DL, DTL, EC, and it's associate 55 could assist to mitigate any pandemics. The possible roles and challenges of these techniques in a pandemic, especially for COVID-19, are mentioned in section 4. This section has tried to bring an overview and general progress of DL, DTL, and EC in the following three subsections. 60 Deep learning (DL) (also known as hierarchical learning or deep structured learning) is one of the great inventions for modern-era of Artificial Intelligence (AI) [11] . Until the decade '90s, classical machine learning techniques were used for making inferences on data and prediction. Nevertheless it had several drawbacks such as depend on handcrafted features, bounded by 65 human-level accuracy, etc [18] . But in case DL, handcrafted feature engineering is not required rather features are extracted from data during training. In addition, DL can make more accurate classifications and predictions with the help of innovative algorithms, computing power of modern machines, and the availability of Big Datasets [19] . Nowadays, DL methods have been successfully applied for several AI-based medical applications such as Magnetic Resonance Imaging (MRI) images analysis for cancer and diabetes diagnoses, conjunction with biometric characteristics, etc [20] . DL is a kind of learning algorithm or model under the umbrella of AI which is based on Artificial Neural Networks(ANN) [21] . These models are 75 trained using dataset through backpropagation algorithm [22] and a suitable optimizer method [23] . The inherent capacities of such DL model such massive parallelism, non-linearity, and capabilities of feature extraction have made them powerful and widely used [19] . There are several variety of DL algorithm such as Convolutional Neural Networks(CNN) [24, 25] , Recurrent 80 Neural Networks (RNN) [26] , Long Short Term Memory(LSTM) [27] , GAN [28] , etc. After success of a CNN-based model, called AlexNet [29] , many deep learning model has proposed such ZFNet [30] , VGGNet [31] , GooglNet [32] , ResNet [33] , DenseNet [34] , etc specially for computer vision tasks [35] . In figure 1 we try to illustrate a typical methodology of a DL based screening 85 system, where the system uses a DL algorithm (CNN) to predict whether the X-ray images of suspected patient's lung is normal or having viral pneumonia or COVID-19 pneumonia. In the time of the COVID-19 crisis, when the numbers of infected patients are at a time very high and the disease is still spreading, many research groups 90 are using the DL techniques for screening COVID-19 patients by detection fever temperature, viral and COVID-19 pneumonia, etc. In addition, DL is using or could be used for other purposes such as patient care, detection systematic social distancing violation, etc [13] . As for reference, S. Wang et.al used a CNN based DL for screening COVID-19 patients with an accuracy, 95 sensitivity, and specificity of 89.5%, 87%, and 88% respectively by using their computed tomography (CT) images [36] . Similarly, in another study [37] L. Wang et.al used chest X-ray images for a screening of COVID-19 cases with 83.5% accuracy. The description of such works is in section 4.1.1. 100 Transfer Learning is a technique that effectively uses knowledge of an already learned model to solve another new task (possibly related or little related) with require of minimal re-training or fine-tuning [38, 39] . Since DL requires a massive training data compared to traditional machine learning methods. So, the requirement of a large amount of labeled data is a big prob- 105 lem in solving some critical domain-specific tasks, specifically the applications for the medical domain, where the making of large-scale, high-quality annotated medical datasets is very complex, and expensive [40] . In addition, the usual DL model requires large computing power such as GPU enabled sever, although researchers are trying hard to optimizing it [41, 42] . Therefore, 110 Deep Transfer Learning (DTL), a DL based Transfer Learning try to overcome this problem [43] . DTL significantly reduces the demand for training data and training time for a target domain-specific task by choosing a pretrained model (trained on another large dataset of same target domain) for a fixed feature extractor [44] or for further fine-tuning [45] . strating the main steps methodology of a Deep Transfer Learning approach, where an untrained model is trained using benchmark dataset (task-1) for feature extraction. Then that pre-trained model is further used to tackle a new challenge such as the task (task-2) of COVID-19 by just replacing only few last layers in the head of the architecture and required fine-tuning. 120 So far, many DTL models have been proposed [14] . A few recent are mentioned reported and discussed in the article. In a research study [43] , Mingsheng Long et.al proposed a joint adaptation network. It learns a transfer network by aligning the joint distributions of multiple domain-specific layers across domains based on a joint maximum mean discrepancy. In another 125 study [46] , Yuqing Gao and Khalid M. Mosalam proposed a state-of-the-art transfer learning model based on VGG model [31] . They have used ImagNet [47] dataset for features extractor and their hand label construction images for fine-tuning. Abnormality classification in MR images through DTL proposed in a study [48] . The authors of that study also have used pre-trained ResNet34 model with fine-tuning. In a research practice [49] , a DTL for diagnosing faults in target applications without labeling was proposed. Their framework used condition distribution adaptation. Q-TRANSFER [50] , another DTL framework proposed by Trung V. Phan et.al. To mitigate the dataset insufficiency problem in the context of communication networking, a 135 DTL-based reinforcement learning approach is used. As the COVID-19 disease spread is terrifying all over the world, screening, quarantine, and providing appropriate treatment to COVID-19 patients has become the first priority in the current scenario. But the global standard diagnostic pathogenic laboratory testing is massive time consuming and more 140 costly with significant false-negative results [51] . At the same time, tests are are hardly to be taken place in the common healthcare centres or hospital due to limited resources and places compared with the high volume of cases at one time. To combat this kind of situation, the researcher from this domain are trying hard to develop some possible DTL models to mitigate this challenges 145 [52, 53] . As for example M. Loey et.al in [52] use DTL along with the GAN model on their very limited, only 307 chest X-ray images to test COVID-19 disease based patient chest X-ray. Here, they have three pre-trained stateof-the-art model namely Alexnet [29] , GoogleNet [32] , and ResNet18 [33] . Among these three pre-trained GoogleNet give the highest accuracy in their 150 experimental studies. In the era of cloud computing, maximum tech companies in the world rely on very few selected cloud providers for hosting and computation power. The user's data from millions of devices around the world is being delivered 155 to some centralized cloud providers for processing or computation. This data transformation always resulted in extra latency and extra bandwidth consumption [54] . The explosive proliferation of IoT devices along with the requirement of real-time computing power have forced to move the scenario of computing paradigm towards Edge computing. Therefore, instead of relying 160 on doing all the work at a cloud, it focuses to start the computational process close to the IoT devices (near to the source of data) in order to reduce the utilized bandwidth and latency [55, 56] . Sometime in Edge computing, an additional nearby server called Fog is associated between the cloud and the Edge or IoT devices. It locally stores the copy of densely used data from the 165 cloud and it provides additional functionality to IoT devices to analyze and process their data locally with real-time working capability. Hence only the relevant data from IoT devices is need to transferred to the cloud through the Fog server [57] . In figure 3 the hierarchy of a possible framework for Edge Computing as-170 sociation with Fog and Cloud computing is illustrated. The data are collected from various IoT devices are being pre-processed before sending by Edge to Fog server for the analysis and computation with the real-time speed (because of the minimal distance between Edge layer and IoT devices and the local database of Fog). While the cloud holds the central control system and 175 it manages the whole database of the system. The database on the cloud is continuously uploaded by the Fog only when it has important data or information. Although EC is not a new concept but it becomes popular in the last five years in the era of IoT [58, 59] . Few recent different type of state-of-the-art of EC are mentioned in this section as a way to familiarize the reader with the recent development with the era of EC and its potential benefits in mitigating COVID-19 as pandemic. EdgeIoT [60] , a study of mobile EC proposed by X. Sun et.al. It is a SDN-based EC work with Fog Computing(FC) [61] to provide computational load locally. In a study [62] , F. Wang et.al have proposed 185 a joint offloading strategy of mobile EC and wireless power transfer. This scheme tried to address energy consumption, latency, and access point issues in IoT. In another study [63] , Wei Ding et.al propose a field-programmable gate array-based depth-wise separable CNN accelerator to improve the system throughput and performance. They have used double-buffering-based 190 memory channels to handle the data-flow between adjacent layers for mobile EC. On the other hand, G. Premsankar et.al in their case study [64] have discussed how efficient mobile gaming can run through EC. In a study [65] , S. Wang et.al have proposed a mobile edge computing with an edge server placement strategy. In their multi-objective constraint optimization-based 195 EC have tried reduced delay between a mobile user and an edge server. In-Edge AI [65] , an integrate the deep reinforcement learning techniques and Federated Learning framework with mobile edge systems are proposed by X. Wang et.al. This framework intelligently utilizes the collaboration among devices and edge nodes to exchange the learning parameters for betterment. In another recent study [66] , an integrated two key technologies, ETSI and 3GPP are introduced to enhanced slicing capabilities to the edge of the 5G network. In the case of COVID-19 like pandemics, discussion of the possible role of EC is done in section 4.3. Although the whole article is referred and cited current relevant stateof-the-art wherever relevant, this section is dedicated to provide a review on some of the very generic recent state-of-the-art works related to transfer learning approaches over edge computing. As mentioned in section 2.1, the progress of DL is very fast but when it comes to application in Edge or 210 IoT devices then a huge gap is noticeable [67] . However, researchers are working hard to cope with the challenges, as results in many computing ideas, optimized model, as well as some computing accelerator devices, comes in picture [68, 69] . Deep Transfer Learning as mentioned in section 2.2 is one such area that is useful where the size of datasets is not sufficient [43] . This transfer learning is also useful where computing resources are not sufficient such as Edge or IoT devices [70] . Since edge computing becomes popular in the last few years, so, we restricted this review to the last five years with chronological order. Lorenzo Valerio et.al have studied the trade-off between accuracy and 220 traffic load of computing in edge-based on transfer learning [71] . They have suggested that sometimes the partial model needs to move across edge devices and data will stay at those edge devices and vice-versa. In a study [72] , [75] . In their data-driven cooperative task allocation scheme, they have used the concepts of the Knapsack problem to prioritized the tasks before transferring them for use in another task. In a study [76] As mentioned in section 1, the appropriate drugs, vaccines, infrastructure are not ready up to some stages of any pandemic. Therefore, to cope with challenges existing knowledge, infrastructures, AI-based models could be exploited to mitigate such pandemic. This section tried to bring four insights of 280 the discussion topics and their roles in mitigating pandemics. Each of them is systematically discussed with potentiality with recent state-of-the-art and challenges. issues [86] . Some of them are Testing Sample Classification, Medical Image Understanding, Forecasting, etc [87] . Some recent DL based models have already proposed to cope with pandemics are listed and their main features are highlighted in table 1. This table brings some proposed peer-reviewed as well as few promising pre-print works. Table 1 has placed some recent works Deep Bayes-SqueezeNet based diagnosis of COVID-19 from X-ray images. • Develop an intelligent diagnosis system for COVID-19 using practical DL networks for medical image processing. • A new decision-making system for COVID-19 with the integration of conventional and state-of-the-art methods for chest X-ray images. C.Butt et.al [89] Screen coronavirus disease 2019 pneumonia. • COVID-19 Patterns Detection in X-ray Images. • Identification of COVID-19 disease. • A resource efficient model with overall accuracy of 91.4%, COVID-19, sensitivity of 90% and positive prediction of 100% in the dataset from [94] . M. Zhou et.al [95] Differentiating novel coronavirus and influenza pneumonias. • An early diagnosis tool on chest CT images for differentiate Coronovirus pneumonia and normal Influenza with transferability. A. Lopez-Rincon et.al [96] Identification of SARS-CoV-2 from Viral Genome Sequences. • Interaction between viromics and DL. • A DL-based model to develop an assisted detection tests for SARS-CoV-2 O.Gozes et.al [97] Automated Detection & Patient Monitoring using Deep Learning and CT Image Analysis. • A model utilizing 2D and 3D DL for clinical understanding. • Proposed a systematic continuous monitoring COVID-19 patients and their clinical data to make a statistical Corona score for monitoring their progress. S.M. Ayyoubzadeh et.al [98] Predict the incidence of COVID-19 in Iran. • Data were mined from Google Trends website • Linear regression and LSTM models were used to estimate positive COVID-19 cases L. li et.al [99] Fully automatic framework to detect COVID-19 using CT images of chest. • Developed a DL model, COVNet to detect COVID-19 by extracted visual features from chest CT exams. • Collected dataset consisted of 4356 chest CT exams from 3,322 patients from six hospital. L. Wang et.al [37] Open source Chest X-Ray Image dataset and a deep CNN for Detection of COVID-19 Cases. • Proposed a publicly available COVID-Net, a deep CNN for the detection of COVID-19 cases from CXR images. • COVIDx, an open access chest X-ray(CSR) dataset consisting 13,800 CSR images across 13,725 patient. S.J. Fong [100] A forecasting model of COVID-19. • A Composite Monte-Carlo simulation forecasting model. • A case study of using above simulation through deep learning. S. Chae [101] Predicting infectious disease using DL and Big data. • A study on DL and LSTM model over ARIMA model to predict future infectious diseases. • Proposed model tried to improve existing surveillance systems to detect future infectious diseases. From table 1 it could be drawn one conclusion that the majority of the works are for assisting radiologists to diagnose diseases. Some of are mentioned forecasting, fake news alert, etc, but more critical parts of this pan-300 demic maybe are addressed by this DL approach. Successfully apply DL in COVID-19 or any running pandemic has three main challenges. The first one is a shortage of reliable datasets. As data collection and validation are a timeconsuming process as well as privacy issues also there whereas a pandemic or epidemic comes suddenly. The second one is the variety of data of a pan-305 demic virus. This COVID-19 virus 'SARS-CoV-2' has mutating itself over different geographic regions, environments, and time [102, 103] . Therefore, the pandemic dataset collected from one region may not be work to drawn inference on the pandemic of other regions. The third one high computational resources required for a DL model whereas to cope with an outbreak 310 IoT or Edge Device (ED) are useful for many purposes [15] . Though these types of equipments have low computing resources. In order to overcome such challenges, cleaver implementation of relevant AI strategies is required. For the first two challenges, DTL and GAN [28] could be a possible approach towards possible solutions. DTL has described 315 in section 4.2 whereas details about GAN are out of the scope of this article. The third challenge could be mitigated using Cloud Computing, Fog Computing, and Edge Computing [104] . However, for Cloud or even Fog Computing latency and data security & privacy could be a problem. Therefore, Edge Computing could be effective for the third challenge, which has 320 described in section 4.3. Section 2.2 has described about Deep Transfer Learning (DTL) in general. In this sections, how DTL could help to mitigate COVID-19 like pandemics 325 is described. As mentioned, sufficient datasets of COVID-19 or any running pandemic are difficult to develop in a short period of time. Therefore, to exploit the benefit of DL to cope with COVID-19 or other pandemics are a bit challenging. Therefore, DTL could be effective in this case. As through DTL a DL model could be trained using a large scale benchmark dataset and 330 learned features could be used in the domain of COVID-19 [53] . Many researchers are trying hard to use this DTL in the domain COVID-19 for many purposes. We have tried to summarize in table 2 some of the recent state-ofthe-art along with their main contribution towards mitigation of pandemics. As the number of peer-reviewed work is limited as this pandemic is new, so this table also has listed some pre-print works, which have tried to introduce some of the contributions in mitigating this current pandemic. • DTL on a subset of 2000 of 5000 radiograms was used to train four popular CNN, including ResNet18, ResNet50, SqueezeNet, and DenseNet-121, to identify COVID-19 disease. • Evaluated these trained models on the remaining 3000 radiograms and achieved a sensitivity rate of 97%( 5%), while a specificity rate of 90% approx. S.Basu et.al [106] Screening COVID-19 using Chest X-Ray Images. • A domain extension transfer learning with pre-trained deep CNN is tuned for classifying among four classes: normal, other diseases, pneumonia, and Covid19. • A 5-fold cross-validation has experimented and overall accuracy measured 95.3%0.02. N.E.M Khalifa et.al [107] An Experimental Case on a limited COVID-19 chest X-Ray dataset. • A study on neutrosophic and deep transfer learning models on limited COVID-19 chest X-Ray dataset. • They first converted grayscale X-ray images into neutrosophic images then applied pre-trained Alexnet, Googlenet, and Restnet18 to classify four classes: COVID-19, Normal, Pneumonia bacterial, and Pneumonia virus. B.R. Beck et.al [108] Predicting commercially available antiviral drugs that may act on SARS-CoV-2 through the DL model and a drug-target interaction. • Drug-target interaction model called MT-DTI to recognize commercially available drugs that could act on SARS-CoV-2. • Proposed a list of antiviral drugs identified by this MT-DTI model. A. Narin et.al [109] Automatic Detection of COVID-19. • Three different CNN (ResNet50, InceptionV3, and Inception-ResNetV2)-based models for the detection of COVID-19 pneumonia infection using X-ray radiography. • Proposed pre-trained ResNet50 has given the best result among these three. I.D. Apostolopoulos et.al [53] Performance evaluation of state-of-the-art CNN architectures through TL over medical image classification. • Suggested DL with X-ray imaging may extract significant bio-markers related to the COVID-19 disease. • A dataset of 1427 X-ray images consisting of 224 images of Covid-19 disease, 700 images of common bacterial pneumonia, and 504 images of normal conditions. B. Subirana et.al [110] New crowdsource AI approach to support health care dealing with COVID-19. • A proposed label works on recognition of cough sound recording by phone as a diagnostic test for COVID-19. N.E.M. Khalifa et.al [111] Detection COVID-19 using GAN and TL method. • A combination of GAN and DTL models for enhancing testing accuracy. • Their ResNet18-based combined model achieved stateof-the-art accuracy in a chest x-ray dataset. DTL does task adaption that is very necessary for analyzing, diagnosing as well as mitigating COVID-19 like pandemics. The number of studies Most of these works could be easier when AI is cooperating and forming such a model along with IoT or ED [13] . Some issues could be solved by EC as described in section 4.3. Though 350 a better system could be delivered when the most suitable algorithm applied on EC. One possibility of archiving this when DTL implemented alongside with EC which as conceptually describes in section 5. Edge or IoT devices-based sophisticated equipments such as smart medical equipment, webcam, drone, wearable sensors, etc are very useful in a pandemic like situations [112] . As mentioned in section 2.3, edge computing brings the computation to near edge devices. It reduces latency, security & privacy issue, etc. Therefore, this computing paradigm will be very effective 360 to mitigate a pandemic situation [113] . The researchers from all over the world are trying hard to bring this along with other AI techniques to mitigate current COVID-19 pandemic [15, 114] . So far only a limited number of studies have investigated the use of EC in obtain an efficient and effective mitigation system of COVID-19. This subsection tried to bring some poten-365 tiality and scopes which shall help to mitigate COVID-19 like pandemics. Table 3 has mentioned some EC based studies on COVID-19 and related healthcare. EC works on site, so, many benefits could draw from EC with IoT or 370 ED. Nevertheless as mentioned IoT or ED has limited computing resources. Therefore, to get the benefit of modern AI algorithm such as DL it is still challenging. To cope with these challenges researchers from all over the world are working hard to propose many ideas [68, 121, 17] . But so far only a few studies on EC in pandemic are proposed in limited areas of application as 375 mentioned in table 3. Assisting many critical COVID-19 related tasks such as remote sensing-based COVID-19 patient monitoring, Hygienic practice monitoring, systematic social distancing monitoring in a crowded area, etc could be done through EC [13] . This article brings a conceptual model of EC with DTL in section 5 as a future scope to cope with such challenges. [115] An open-source EC for clinical screening system. • Fever and Cyanosis detection using visible and farinfrared cameras emergency department. • This image segmentation-based EC uses open source. hardwares. A.A.Abdellatif et.al [116] Data and applicationspecific energy-efficient smart health systems • An optimizes medical data transmission from edge nodes to the healthcare provider with energy efficiency and quality-of-service. • Managing a heterogeneous wireless network through EC to provide fast emergency response. A.H. Sudhro et.al [117] QoS optimization in medical healthcare applications. • A window-based Rate Control Algorithm to QoS in mobile EC. • A framework for Mobile EC based Medical Applications. M.Chen et.al [118] Smart Healthcare System. • Edge cognitive computing-based smart healthcare mechanism to dynamic resource allocation in healthcare. P.Pace et.al [119] Efficient Applications for Healthcare Industry 4.0 • Proposed BodyEdge, an architecture suited for humancentric applications in context of the emerging healthcare industry. • A tiny mobile client module with EC for better health service. H.Zhang et.al [120] Smart Hospitals Using Narrowband-IoT. •An architecture to connect intelligent things in smart hospitals based on Narrowband IoT. • Smart hospital by connecting intelligent with low latency. Data is the fuel of a modern computing. Whether it is medical field or retailer market, in every field data are the most precious things. Recent AI techniques are mostly follow data driven approaches [122, 123] . DL or DTL 385 based algorithms almost fully depend on the dataset. Therefore, to cope with a pandemic, data is one of the driving forces. For a pandemic as COVID-19, the dataset could be chest X-ray, CT images, pathological images, geographical region based spreading patterns, seasonal behavior, regional mortality rates, impact on the economy, etc. [124] . In table 4 some available datasets 390 related to COVID-19 like pandemics are mentioned with brief description. As mentioned data is driving force to which bring the knowledge but it not easily available. Specially COVID-19 or a sudden pandemic, gathering data and arrange it in a knowledgeable form are not expected as an easy 395 task. Although for COVID-19, many sectors are very active as a result many data sources are quickly oriented towards COVID-19 pandemic. Some data Table 4 : Some Dataset for COVID-19 like pandemics. Brief description COVID-CT-Dataset [125] . • A publicly CT scan dataset consisting of 275 positives for COVID-19 cases. COVID-19 X-ray image dataset with two different combinations for Applying with DTL models for experimenting with different experimental setup. [53] • One dataset of 1427 X-ray images consisting of 224 images of Covid-19 disease, 700 images of common bacterial pneumonia, and 504 images of normal conditions. • Another dataset of 1442 X-ray images consisting 224 images of Covid-19 disease, 714 images of common bacterial pneumonia, and 504 images of normal conditions COVID-19 Image Data Collection [94] . • It was a crowdsourcing hosting currently contains 123 frontal Xrays images. Chest CT Images [99] • A dataset consisted of 4356 chest CT exams from 3,322 patients. • Data are collected from six hospitals of average age is 4915 years, among them 1838 male patients. Coronavirus Twitter Dataset [126] . • A multilingual COVID-19 Twitter dataset that has been continuously collecting since Jan 22, 2020. • It consists online conversation about COVID-19 to track scientific misinformation, rumors, etc. COVIDx CXR Dataset [37] . • This dataset consisting of 13,800 images of chest radiography across 13,725 patients. Epidemiological COVID-19 data [127] . • Individual-level data from municipal, provincial, and national health reports, as well as additional information from online reports. • All data are geo-coded including where available, including symptoms, key dates, and travel history. H1N1 Fever Dataset [128] . • Two datasets collected at Narita International Airport during the H1N1 pandemic 2009. • The first dataset only 16 candidates and the second one is 1049 collected using infrared thermal scanners. Registry data from the 191820 pandemic. [129] . • A high-quality vital registration data with mortality for the 191820 pandemic from all countries. sources are listed in table 5 where COVID-19, as well as other pandemic data, are available, so, researchers may use them for many purposes. The main challenges are sufficient datasets especially machine-readable datasets 400 in every affected sector are yet to be available. Therefore, that are the challenges for data-driven AL algorithms or models, hence existing studies on real data and analysis are few. Although some datasets mentioned in table 4 but most of them are for clinical purposes. As said this novel coronavirus is behaving differently across geographic regions, different environments, etc. Therefore, data of one region may not be effective to enhance knowledge in other regions. Data privacy and security also are considered ones of the big issues. To this reason this article suggesting transfer learning approaches to be used in developing models for mitigating COVID-19 like pandemics or epidemics. [3] • WHO leading this battle by providing each and every possible data and information. • Most of the data are unstructured so it bit challenging to feed into an AI model. Johns Hopkins University is in the forefront to provide COVID-19 dataset [130] through their portal: • An open repository where many datasets is stored. • Many research projects stores their data and mentioned links to their article, but they provide a link to see and access the COVID-19 dataset. Kaggle: https://www.kaggle.com/c/covid19global-forecasting-week-#. • An online community of data scientists and machine learning practitioners • Forecasting dataset and other COVID-19, or pandemic dataset available. As mentioned 4.1.2, DL has some limitations to cope with the challenges of a pandemic whereas section 2.2 has described the task adaptability through DTL where data shortages are there. Section 2.3 mentioned the potentialities of EC where computing power is low. Therefore, the merging of these 415 three computing models could be more effective in assisting the mitigation of pandemic situations. This combined model, that is, Deep Transfer Learning over Edge Computing(DTLEC) will take the power of DL through DTL as well as would be applicable in critical sectors by EC to cope with a sudden pandemic. There are some studies that exist in DTLEC as in [68] and some 420 related work mentioned in section 3. However, these works are still in general concept or their proposed methods are applicable only to some others application areas. As per literature studies, this idea has not been studied or experimented to mitigate COVID-19 pandemic. This section tried to present a precedent working pipeline of DTLEC to assist mitigation of pandemic or 425 epidemic. DTLEC model could be helpful in the healthcare sector, quarantine center, or other critical areas where an outbreak may arise. As in figure 4 edge or IoT devices that are set up in those areas may be embedded with EC, 430 and then it could be connected with a cloud server. A state-of-the-art DL model shall train in GPU enabled cloud server by using a benchmark dataset for feature extraction. Then a pre-trained model(with extracted weight or features except for classification layer) shall push down to the edge devices. In edge devices required fine-tuning mechanism to be implemented into that 435 model with some real data. In this way, the model may ready to work in some critical areas where outbreaks are affected such as hospitals, crowded places, and many more. In figure 5 a typical current COVID-19 outbreaks situation and possible working model are shown. This figure illustrating the proposed framework 440 to tackle the COVID-19 situation by using DTL in EC in both COVID-19 patient care and management systematic social distancing. In the first scenario, we may use several healthcare sensors like blood pressure sensors, body temperature sensors, webcam, etc to sense the data about the running health condition of each patient. Then all of the collected data would be 445 sent to the EC layer where a pre-trained DL based model will be used to process the captured data and making an inference out of it. If the generated report is a critical health condition then an automatic system alert message will be sent with all the details to the hospital control room and also to all the doctors of associated team. In the second scenario, several public place 450 monitoring sensors (like a drone, CCTV, traffic cameras, etc) could be used to detect unnecessary illegal crowd with or without wearing masks with help from the DTLEC-based model. If the model finds any such gathering then an automatic system alert message will be sent with all the details to the nearby police station. 455 This model may be successfully deployed in some critical sectors such as hospitals, airports, markets, emergency service areas, and those areas which are the primary hotshots for spreading pandemics. The model has to be work on real data to draw the inference. In order to make it successfully 460 deployed, lots of collaborative work need to be done, which may face many challenges. Some challenges need to be addressed such as (i.) At first, IoT or Edge devices need to be connected with each other and a cloud server, hence a modified sensor type of networking protocol will be required. (ii.) EC through DTL need be implemented, for that appropriate pre-trained deep learning need be carefully selected after some studies. (iii.) For the transfer learning approach, using only EC is not sufficient, while the adoption of EC-Fog-Cloud combined model would be more useful. A deep learning model shall be trained at a cloud server using a benchmark dataset for feature extraction. After that pre-trained model will push down to the edge where 470 limited re-training (or fine-tuning) shall be carried out to orient a few last layers for required inference. So, at least a small pandemic dataset needs to be created. Here, the Fog server could work as a cluster. (iv.) Security and privacy issues of data need to be addressed. This inquires much more attention by researchers in analyzing numerous vulnerabilities that associated 475 with such outbreak due to rumours and fake news. Besides, the privacy of captured data from multiple sources (things in IoT or individuals) will open a new research direction for the near coming future. (v.) A new simulation model may be required for experimental studies. These are a few of the other many challenges we can work for. 480 This article has tried to bring potentialities and challenges of Deep Transfer Learning, Edge Computing and their related issues to mitigate COVID-19 pandemic. It has also proposed a conceptual combined model with its scope and future challenges of working at critical sites and real data. As the running 485 pandemic is very new, so, there is a limited number of peer-reviewed studies and experimental results. Therefore, this systematic study article also considered some pre-print studies which are tried to make some contributions in mitigating running pandemic. The running pandemic definitely will be mitigated but there will be a left over impact on global health, economics, 490 education, etc, so mitigation of this pandemic is necessary to restrict further worsen. Every scientific community of the world needs to think wisely to get prepared cope with such kind of crisis in a case similar outbreaks appear in the future. This article will definitely assist to the research community; especially deep transfer learning and edge computing to work further 495 in developing many tools and applications towards the mitigation of running pandemic or any future pandemic if that could arise. The epidemiology and pathogenesis of coronavirus disease (covid-19) outbreak Virology, epidemiology, pathogenesis, and control of covid-19 Coronavirus disease 2019 (covid-19): situation report World health organization declares global emergency: A review of the Influenza pandemics: past, present and future Rational use of personal protective equipment for coronavirus disease ( covid-19) and considerations during severe shortages: interim guidance Covid-19navigating the uncharted Analysis and forecast of covid-19 spreading in china, italy and france Deep learning Review of artificial intelligence techniques in imaging data acquisition, segmentation and diagnosis for covid-19 Insights of artificial intelligence to stop spread of covid-19 A survey on deep transfer learning Digital technology and covid-19 Edge computing: A survey Convergence of edge computing and deep learning: A comprehensive survey, IEEE Communications Surveys & Tutorials Handcrafted vs. non-handcrafted features for computer vision classification Deep learning A survey on deep learning in medical image analysis Evolving artificial neural networks Learning internal representations by error propagation The marginal value of adaptive gradient methods in machine learning Convolutional networks for images, speech, and time series Fundamental concepts of convolutional neural network, in: Recent Trends and Advances in Artificial Intelligence and Internet of Things Recurrent neural networks for prediction: learning algorithms, architectures and stability Long short-term memory recurrent neural network architectures for large scale acoustic modeling Advances in neural information processing systems Imagenet classification with deep convolutional neural networks Visualizing and understanding convolutional networks Very deep convolutional networks for large-scale image recognition Going deeper with convolutions Deep residual learning for image recognition Densely connected convolutional networks Advancements in image classification using convolutional neural network A deep learning algorithm using ct images to screen for corona virus disease Covid-net: A tailored deep convolutional neural network design for detection of covid-19 cases from chest radiography images A survey on transfer learning Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Artificial intelligence (ai) systems for interpreting complex medical datasets deep learning hardware: Past, present, and future A survey of techniques for optimizing deep learning on gpus Deep transfer learning with joint adaptation networks Traditional feature engineering and deep learning approaches at medical classification task of imageclef 2016 An ensemble of fine-tuned convolutional neural networks for medical image classification Deep transfer learning for image-based structural damage recognition 2009 IEEE Conference on Computer Vision and Pattern Recognition Application of deep transfer learning for automated brain abnormality classification using mr images Deep transfer network with joint distribution adaptation: A new intelligent fault diagnosis framework for industry application Bauschert, q-transfer: A novel framework for efficient deep transfer learning in networking A review of salivary diagnostics and its potential implication in detection of covid-19 Within the lack of covid-19 benchmark dataset: A novel gan with deep transfer learning for corona-virus detection in chest x-ray images Covid-19: automatic detection from x-ray images utilizing transfer learning with convolutional neural networks Serverless computing: Current trends and open problems The emergence of edge computing Introducing the new paradigm of social dispersed computing: Applications, technologies and challenges A survey and analysis of security threats and challenges Edge computing: Vision and challenges All one needs to know about fog computing and related edge computing paradigms: A complete survey Edgeiot: Mobile edge computing for the internet of things A survey of fog computing: concepts, applications and issues Joint offloading and computing optimization in wireless powered mobile-edge computing systems Designing efficient accelerator of depthwise separable convolutional neural network on fpga Edge computing for the internet of things: A case study Edge server placement in mobile edge computing Toward slicing-enabled multi-access edge computing in 5g Deep learning with edge computing: A review Learning iot in edge: Deep learning for the internet of things with edge computing A survey on optimized implementation of deep learning models on the nvidia jetson platform On-edge multi-task transfer learning: Model and practice with data-driven task allocation Accuracy vs. traffic trade-off of learning iot data patterns at the edge with hypothesis transfer learning Proactive content caching by exploiting transfer learning for mobile edge computing Bandwidth-efficient live video analytics for drones via edge computing Are existing knowledge transfer techniques effective for deep learning with edge devices? 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS) Ai-enhanced offloading in edge computing: when machine learning meets industrial iot Pruning deep convolutional neural networks for efficient edge computing in condition assessment of infrastructures Cartel: A system for collaborative transfer learning at the edge Fedhealth: A federated transfer learning framework for wearable healthcare Openei: An open framework for edge intelligence Deep learning for ultra-reliable and low-latency communications in 6g networks A new load balancing strategy by task allocation in edge computing based on intermediary nodes Urban intelligence with deep edges Mobileda: Towards edge domain adaptation Resource usage and performance trade-offs for machine learning models in smart environments Deep learninga technology with the potential to transform health care Deep learning intervention for health care challenges: Some biomedical domain considerations Covidiagnosis-net: Deep bayes-squeezenet based diagnostic of the coronavirus disease 2019 (covid-19) from x-ray images Deep learning system to screen coronavirus disease 2019 pneumonia Weakly supervised deep learning for covid-19 infection detection and classification from ct images Automated detection of covid-19 cases using deep neural networks with x-ray images Yolov3: An incremental improvement Towards an efficient deep learning model for covid-19 patterns detection in x-ray images Covid-19 image data collection Improved deep learning model for differentiating novel coronavirus pneumonia and influenza pneumonia Accurate identification of sars-cov-2 from viral genome sequences using deep learning Rapid ai development cycle for the coronavirus (covid-19) pandemic: Initial results for automated detection & patient monitoring using deep learning ct image analysis Predicting covid-19 incidence through analysis of google trends data in iran: Data mining and deep learning pilot study Composite monte carlo decision making under high uncertainty of novel coronavirus epidemic using hybridized deep learning and fuzzy rule induction Predicting infectious disease using deep learning and big data Genomic diversity of sars-cov-2 in coronavirus disease 2019 patients We shouldnt worry when a virus mutates during disease outbreaks Deep-covid: Predicting covid-19 from chest x-ray images using deep transfer learning Deep learning for screening covid-19 using chest x-ray images A study of the neutrosophic set significance on deep transfer learning models: An experimental case on a limited covid-19 chest x-ray dataset Predicting commercially available antiviral drugs that may act on the novel coronavirus (sars-cov-2) through a drugtarget interaction deep learning model, Computational and structural biotechnology journal Automatic detection of coronavirus disease (covid-19) using x-ray images and deep convolutional neural networks Hi sigma, do i have the coronavirus?: Call for a new artificial intelligence approach to support health care professionals dealing with the covid-19 pandemic Detection of coronavirus (covid-19) associated pneumonia based on generative adversarial networks and a fine-tuned deep transfer learning model using chest x-ray dataset Iot pcr for pandemic disease detection and its spread monitoring An agent-based self-organizing model for large-scale biosurveillance systems using mobile edge computing, Simulation Modelling Practice and Theory Cognitive computing and wireless communications on the edge for healthcare service robots Autotriage-an open source edge computing raspberry pi-based clinical screening system Edge computing for energy-efficient smart health systems: Data and application-specific approaches Mobile edge computing based qos optimization in medical healthcare applications Edge cognitive computing based smart healthcare system An edge-based architecture to support efficient applications for healthcare industry 4.0 Connecting intelligent things in smart hospitals using nb-iot Deep learning for iot big data and streaming analytics: A survey Data driven feature selection for machine learning algorithms in computer vision A survey of data-driven and knowledge-aware explainable ai Open data resources for fighting covid-19 Covid-ct-dataset: A ct scan dataset about covid-19 Covid-19: The first public coronavirus twitter dataset Epidemiological data from the covid-19 outbreak, real-time case information Fever screening during the influenza (h1n1-2009) pandemic at narita international airport, japan Estimation of potential global pandemic influenza mortality on the basis of vital registry data from the 1918-20 pandemic: a quantitative analysis A county-level dataset for informing the united states' response to covid-19 None