With the onset of the Covid-19 pandemic, there has been a tremendous impact on the lives of people globally. The global tally of the no. of infected cases is 229,293,200 and the total death toll is 4,705,498 as of September 20, 2021, and this is just the no. of accounted cases. Apart from this, the covid-19 genome sequence is continuously mutating which resulted in the generation of different and more dangerous types of variants like the delta and delta plus to name a few. The globe witnessed the horrific scenes caused due to the shortage of resources in terms of healthcare during the unprecedented first and second waves. The leading scientists have already predicted the onset of the third wave which is expected to start in the last quarter of the year 2021.

To tackle the third wave, the governments have started the vaccination campaign for the people. But still, as per the earlier predictions the third wave is inevitable and the third world countries that are lagging in terms of medical infrastructure would be affected the most in the oncoming third wave. Reverse Transcription-Polymerase Chain Reaction (RT-PCR) is considered as the standard reference for the Covid-19 based testing but it has certain disadvantages like higher testing costs and longer processing time for the collected samples. Considering the factors like lack of essential infrastructure required for the testing and the capital required, we are hereby proposing a novel approach as a preparatory measure for detecting the presence of Covid-19 using Lung Ultrasound Imaging which in turn can be further deployed on a real-time basis with the help of Intel FPGA Cloud Connectivity Kit and Azure IoT Suite.

Compared to CT scans and X-Ray based detections, Ultrasound Imaging is low cost and radiation-free detection method. In this proposed approach, we would conduct a thorough literature review on the existing deep learning models which has been developed for predicting the covid-19. Based on the earlier approaches, a new model would be proposed which will be extensively trained with available datasets for addressing the limitations offered by the earlier models. Upon the satisfactory performance of the model in terms of accuracy, precision, and computation time, it would be deployed onto the FPGA Cloud Connectivity kit for real-time application. With the help of Azure IoT support, the Ultrasound Images from different centers can be obtained and the test results can be again diverted back to the respective centers at a significantly less amount of time as compared to the RT-PCR test.

This proposed approach will help set up a nationwide/worldwide low-cost testing facility with a significantly less amount of capital being invested. With this approach, we are intending to address the two major challenges prevailing in the present scenario with the first one being high testing costs and improved timing for generating the test results. This approach would in turn be helpful to combat the oncoming third wave of this Covid-19 global pandemic.

Project Proposal

1. High-level project introduction and performance expectation

Coronavirus disease or COVID-19 is caused by the virus named SARS-CoV-2 (severe acute respiratory syndrome coronavirus 2) as named by the International Committee on Taxonomy of Viruses (ICTV). The globe has witnessed the havoc caused by this disease wherein the people were gasping for oxygen, rushing from one hospital to another for getting the admissions. Along with this, many of them have also lost their near and dear ones because of this virus. The healthcare system across the world was on the verge of collapse during the first and second wave of the Covid-19 pandemic and now countries across the globe are witnessing the surge in the daily no. of infections which is indicates the beginning of the third wave of the Covid-19 pandemic.

As per the studies and the observations which were made through the research of the first and second waves of the covid-19 pandemic along with the several historical events of a pandemic similar to this magnitude, it has been observed that personal hygiene, early detection, and vaccination play a major role in curbing the spread. As compared to the developed countries, the third world countries lag in terms of essential medical infrastructure. So, it is of prime importance for containing the spread of infection in these areas else the scale of damage would be much more than that observed earlier. Governments across the world have rolled out the vaccine campaign but still, it will not be enough to stop the oncoming third wave of the pandemic.

With the onset of Machine learning and Artificial Intelligence, and their increasing use cases in the field of the healthcare industry, a feasible solution can be proposed by the deployment of these technologies. Early detection is one of the key factors for suppressing the spread. Reverse Transcription-Polymerase Chain Reaction (RT-PCR) is considered as the standard reference for the Covid-19 based testing but it has certain disadvantages like higher testing costs and longer processing time for the collected samples. With the help of machine learning and deep learning techniques, researchers are working to develop methods and models for detecting the covid-19 virus.

Scientists and researchers are continuously developing and deploying deep learning models and algorithms for the detection of Covid-19 traces. The models are being trained on the datasets comprising of Chest X-Rays, CT-Scans for tracing the Covid-19 virus, and some of those models have attained accuracies of more than 90%. But the problem is that the X-Ray and CT Scan diagnostics cost some amount of fortune and along with that, the patient will be subjected to ionizing radiation. So, for solving the above-mentioned problem, a low-cost yet precise alternative has to be explored. Fortunately, the solution can be explored through the means of Ultrasound diagnostics. Ultrasound diagnostics provide a cost-effective imaging solution as compared to X-Ray and CT- Scan Imaging Techniques and it does not use any kind of ionizing radiation. The Datasets for Lung Ultrasound diagnostics are also available in the public domain so it is possible for model training.

The deep learning Model uses Convolutional Neural Networks (CNNs) for segmenting the images in the form of layers to extract the necessary parameters and features. With the help of these extracted features, the model predicts the nature of the image. Field Programmable Gate Arrays (FPGAs) come into the picture mainly for the implementation of these networks on the hardware platform. FPGAs are also used as hardware accelerators to lower the rendering time taken by the models while in training.

So, with this background, we are hereby proposing FPGA-based deep learning architecture for detecting the Covid-19 Virus using Lung Ultrasound Image which can be further deployed as a testing scheme. This proposed approach will help set up a nationwide/worldwide low-cost testing facility with a significantly less amount of capital being invested. With this approach, we are intending to address the two major challenges prevailing in the present scenario with the first one being high testing costs and improved timing for generating the test results. This approach would in turn be helpful to combat the oncoming third wave of this Covid-19 global pandemic.

2. Block Diagram

3. Expected sustainability results, projected resource savings

The performance of any kind of deep learning model is measured in terms of precision, sensitivity, and specificity. 
Precision is a metric that quantifies the number of correct positive predictions made. Precision, therefore, calculates the accuracy for the minority class. It is calculated as the ratio of correctly predicted positive examples divided by the total number of positive examples that were predicted. Sensitivity is a measure of the proportion of actual positive cases that got predicted as positive (or true positive). 
Sensitivity is also termed Recall. This implies that there will be another proportion of actual positive cases, which would get predicted incorrectly as negative (and, thus, could also be termed as the false negative). This can also be represented in the form of a false negative rate. The sum of sensitivity and false negative rate would be 1.
Specificity is defined as the proportion of actual negatives, which got predicted as the negative (or true negative). This implies that there will be another proportion of actual negative, which got predicted as positive and could be termed as false positives. This proportion could also be called a false positive rate. The sum of specificity and false positive rate would always be 1.
Some models which are using CT scan and X-Ray Image Dataset for computation have achieved precision, sensitivity, and specificity in the range of 90%- 97%. But those models underwent exhaustive training on the high-performance GPU cores and a huge amount of capital was invested for garnering the result. But with regards to this proposed design, Intel FPGA would be used. So along with the above-mentioned three parameters, power consumption also has to be optimized to make the device portable.
So, by considering the above-mentioned parameters, a suitable deep learning model architecture would be implemented on the FPGA kit. This deployed model will be further optimized in terms of resource allocation of the FPGA board, power consumption, rendering speed, etc. For time being, the selected architecture will be trained enough to get the precision, sensitivity, and specificity of more than 65% on the publicly available Lung Ultrasound Datasets.

4. Design Introduction

The Ultrasound Diagnostic is a low-cost, nonionizing radiation-free imaging scheme. The medical fraternity across the globe has the necessary infrastructure available for conducting the diagnostics and no specialized labs or centers are required for setting up. So along with low-cost, ultrasound diagnostics provide an option for implementing a scalable architecture.
So, as the next step, the deep learning model architecture suitable for extracting the features from ultrasound images has to be exhaustively studied and a suitable model has to be selected for the implementation. The model will be trained intensively on the datasets available in the public domain and would be optimized in terms of parameters like precision, sensitivity, and specificity. Then the model will be implemented on the FPGA kit wherein it would be further optimized with regards to resource allocation, power, rendering speed so that the hardware can be deployed on a real-time basis.
Now with the virtue of the Azure IoT Suite, a cloud database can be created. In this database, different centers will upload the lung Ultrasound Image of the patients. The deployed model can be used to extract the features from these images uploaded in the database and after computing, the results can be passed on to the respective centers on short notice. So, this proposed design architecture will help provide an early detection scheme for the covid-19 virus.

5. Functional description and implementation

For machine learning, artificial intelligence, and deep learning applications, many reference models and algorithms have been proposed for segmentation, feature extraction, and pixel-based weighted loss function calculations. The popular reference models and algorithms used across a wide range of domains like healthcare, autonomous driving, agriculture, etc. are Unet, Bonnet, DenseNet, etc. 
Along with these popular models, there are specific models used within the domain of healthcare and biomedical signal and image processing for the detection of pneumonia, breast cancer, tumors, fibroids, kidney stones, etc. So, for this particular project, we will exhaustively study these different models and after garnering the required data, a model suitable for covid-19 diagnostics will be deployed. For this model deployment onto the FPGA board, the HDL Code of the model/algorithm will be developed using the Quartus Prime software. Upon the deployment, it would be subjected to extensive training on the publicly available Covid-19 lung Ultrasound datasets.
For every iteration, the precision, sensitivity, and specificity obtained will be observed and upon satisfactory performance, the parameters like rendering speed, power, and resource utilization of the FPGA device will be optimized. Once the parameters are optimized, the real-time deployment along with the integration of Azure IoT Cloud Suite will be performed.
With the help of the cloud database made available from Azure IoT, the ultrasound diagnostic images can be uploaded on it, and the model will perform the diagnostics on the uploaded images and once the results are obtained, they would be passed on to the respective centers. So, this proposed architecture will work as a low-cost scalable covid-19 diagnostic tool.

6. Performance metrics, performance to expectation

The performance parameters which would be of point major focus on this design would be the precision, sensitivity, and specificity of the proposed deep learning model, along with the rendering speed, power consumption and resources used when deployed onto the FPGA.
The model will be extensively trained to achieve at least a minimum goal of 65% in terms of precision, specificity, and sensitivity. Along with this, it will be ensured that device-level synthesis will be performed for keeping the power consumption, rendering speed in check.  
Intel FPGA is specifically used because the model algorithm which has to develop using HDL Coding can be easily deployed with the help of Quartus Prime Software. The DE-10 cloud connectivity kit has a huge pool of resources in terms of LUTs, DSPs, etc. and along with this, it could be connected with a cloud database with the help of the internet. Because of this, the diagnostics center can upload the ultrasound images on the cloud, and the model after performing the analysis can generate results and send them back to the respective centers. So this eliminates the need for extra hardware at each center in turn proving this solution as a cost-effective one.

7. Sustainability results, resource savings achieved

8. Conclusion