Automatic system for COVID-19 Classification

Diseases ranges from illness to MERS (Middle East Respiratory Syndrome) and SARS (Severe Acute Respiratory Syndrome) are mostly observed due to large family of viruses known as Coronavirus. A new strain of these viruses officially called as COVID (corona virus disease)-19 has originated from Wuhan,

Project Title

Automatic system for COVID-19 Classification

Project Area of Specialization

Artificial Intelligence

Project Summary

Diseases ranges from illness to MERS (Middle East Respiratory Syndrome) and SARS (Severe Acute Respiratory Syndrome) are mostly observed due to large family of viruses known as Coronavirus. A new strain of these viruses officially called as COVID (corona virus disease)-19 has originated from Wuhan, China and spread worldwide. The virus is spread by droplets from an infected person’s mouth when they cough or the droplets can come from their nose when they sneeze. These people generally effected with other people; the virus gets deep into the lungs and causes a severe infection: pneumonia. According to World Health Organization, detection of biological RNA from sputum has a relatively low positive rate in the initial stage to discover COVID-19. The COVID-19 has different characteristics as compared to healthy images, manifested by computer tomography (CT) image. However, doctors’ still requires initial detection of this new pneumonia as early as possible.

The aim of this research is to develop an application based on the deep learning models for classification and localization of the COVID-19 features. This application has two phases. In phase I, initial screening is performed using ResNets-50 model with full learning to classify the COVID-19 affected images and healthy images. In addition, GRAD-CAM model is also presented after classification to visualize the discriminative features for better understanding. In phase II, COVID-19.YOLOv2. ResNets 50 model is used for localization of the infected regions of the patient’s data accurately which will be helpful for the radiologists. The presented application will be helpful to diagnose the disease at an initial stage, which definitely increase the patient survival rate.

Project Objectives

Our major objective of this research study is developing the artificial intelligence (AI) model for localization and the prediction of infected regions.  The chest X-ray and CT images are used as an input for the prediction as follows:

• Automated system is developed for COVID-19 screening and localization based on convolutional neural network and YOLOv2 deep learning model.
• Propose a Deep Convolutional Neural Network (DCNN) to enhance the quality of the CT images.

• Compute best Deep features using rank feature selection approach.
• Accurate classify the healthy and COVID-19 CT images, Healthy vs. COVID-19.
• Propose an automated system based on COVID-19.YOLOv2. ResNets50 models for localization of the affected region in CT images. 
• Improve detection accuracy as compared to the recent deep learning techniques.

Project Implementation Method

AI models play vital role for the detection of diseases because in which highly discriminative features are extracted automatically in the form of hierarchy to learn the more complicated patterns easily. However, AI does not perform well for the detection of pharmaceutical, COVID-19 and epidemiological points of view. The CT images are noisy and much outlier are including in the input date. Recently, no automated system is presented for recognition of COVID-19 in Pakistan. Recent literatures investigate that the transfer learning models are helpful to learn the features of the CT images. The discriminative features selection for accurate detection of COVID-19 using CT images is still a challenging task.

In proposed methodology, the quality of the CT images is enhanced using deep convolutional neural network (DCNN). The enhanced images are passed to ResNets 50 model and best features are selected based on rank feature selection approach. The final optimized features vector passed to different machine learning classifiers for discrimination among the healthy and the COVID-19 images. A model is proposed named as "COVID-19.YOLOv2". ResNets.50 based on the fusion of ResNets 50 and tinyYOLOv2 for accurate localization of COVID-19 features. In this model ReLU-activation 40 layer is selected for features extraction and remaining layers are removed. The extracted features are concatenated with the YOLOv2 layers. The presented model can be utilized in real time applications for COVID-19 detection in less computational time and also minimizes the workload of the radiologists.

The systemic overview of the presented methodology is shown in the Figure 1.

              Figure 1. Overview of proposed method for COVID -19 detection and classification    

The developed application will be embedded into the Ultra96/or any other device. User input the CT images into this device. The device will classify and localize the COVID-19 features in less computational time without taking any help. The overview of the hardware implementation is shown in Figure2.

 Figure 2. Hardware implementation of proposed approaches

Benefits of the Project

Artificial intelligence (AI) models play vital role for the disease detection. Recently, no automated system is presented for recognition of COVID-19 in Pakistan. In the current scenario of COVID-19 pandemic, it is expected to have much increased patient’s burden. As Pakistan is a developing country with limited health care resources, insufficient diagnostic equipment and trained radiologists. So, if we incorporate this artificial intelligence tool in CT chest reporting, it will save time, will mark the diseased part of lungs with more accuracy and thus will minimize the radiologist’s work load. Moreover, it can be used as an initial screening tool along with PCR testing, which will help to identify infected cohort more efficiently and speedily, thus will help to control the spread of disease. Also upon serial scanning of COVID-19 positive ITU patient’s response to the treatment can be monitored within few minutes by utilizing this real time application.

Radiologists as well as customer used this application, simply input the CT images to Utra96 device, This device classify and localize the COVID-19 images. The overview of this project is shown in Figure 3.

                                                       Figure 3. Project overview

Figure 3 shows that overall working of this project. In this project presented deep learning algorithms are embedded into the Ultra96 device. The user manually just pass input image into the Ultra96 device and plug into laptop. This device will mark the COVID-19 features more accurately. 

We have developed the Prototype so far for COVID-19 screening and localization. Few sample screening and localization results are shown in the following Figure 4 for more undemanding of the project. The Prototype will be converted in to real time application with the enhanced results that could be useful in Hospitals if this project proposal will be accepted.

Figure 4: Few sample screening and localization results of proposed algorithm

The developed application accurately screening the COVID-19 and healthy images and also localize the exact pathological region. These cases are confirmed COVID-19 that is also verified through trained radiologists.

                                                                 

Technical Details of Final Deliverable

Software Tools:

• Matlab 2020a/Open CV
• Deep learning toolkit (DLTK)
• Radiant DICOM reader

Project Equipment Details:

• Avnet Ultra96:
  • Part Number: AES-ULTRA96-G,
  • Device Support: Zynq Ultra Scale+ MPSo,
  • Vendor: Avnet
• GPU: 
  • ASUS TUF Gaming GeForce GTX 1650 OC Edition 4GB GDDR6 Video Graphics Card (TUF-GTX1650-O4GD6-P-GAMING)

Our application will contains two  parts, the first part is training, which we will be using different sets of cancer image database to train a machine learning algorithm (model) with their corresponding labels using Graphic Card. The second part is deploying on the edge, which uses the same model we've trained and running it on an Edge device, in this case Movidius Neural Computing Stick through Ultra96 FPGA. This way GPU can run inference while FPGA can do the OpenCV

 

Final Deliverable of the Project

HW/SW integrated system

Core Industry

Medical

Other Industries

Health

Core Technology

Artificial Intelligence(AI)

Other Technologies

Others

Sustainable Development Goals

Good Health and Well-Being for People

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com