Document Type : Primary Research paper
Viveka Shenoy K is student of Computer Science and Engineering Department, PSG College of Technology, Coimbatore, India.
Dr V Maheswaran is Associate Professor, Department of Radiology, PSG IMSR,Coimbatore, India.
Dr G R Karpagam is Professor and Associate Head, Department of Computer Science and Engineering, PSG College of Technology, Coimbatore, India.
Dr B Vinoth Kumar is Associate Professor, Department of Information Technology, PSG College of Technology,Coimbatore, India.
Deep learning has been one of the widely used techniques in biomedical image segmentation. COVID-19 is spreading rapidly all over the world. Early diagnosis of disease is necessary to control the pandemic. The common method of diagnosis of COVID-19 infection is RT-PCR (Reverse Transcription - Polymerase Chain Reaction) method. However, RT-PCR is prone to produce a number of false results and availability of RT-PCR kits is limited. Computed Tomography (CT) and X-Ray of lung can be used as an alternative tool to make the diagnosis, as the disease primarily targets the epithelial cells of the lung. The CT scanners considered to be diagnostic tools for COVID-19. The number of images to be analyzed by the radiologist is large therefore it is necessary to automate the task of CT Image processing. In this paper, we have developed an Automatedbiomedical segmentation system using five different deep learning architectures which performs the COVID-19 lung infection segmentation. The models are trained using open source datasets. Five deep learning models selected for study are U-Net, LinkNet, ResU-Net, ResU-Net++ and U-Net++. The U-Net++ model shown better results with a dice coefficient of 84.69%, sensitivity of 78.92%, specificity of 99.51% and precision of 92.12%. Experimental results are generated using PSG IMSR (Institute of Medical Sciences & Research) dataset, which is a collection of 75,000 images of nearly 125 patients. The radiologists have verified the results of infection segmentation. Automated infection segmentation of HRCT scans using deep learning models can be used for faster diagnosis of COVID-19.
Impact Statement — RT-PCR method used for COVID-19 infection diagnosis is prone to produce a number of false results and the availability of a number of RT-PCR testing kits is limited. Medical imaging like chest CT scan can be used as alternative tool as the disease primarily targets lung. Due to wide availability of CT scanners, they are considered to be diagnosis of COVID-19. The number of patients who undergo scans and images to be analyzed by the radiologist is very large therefore it is necessary to automate the task. Deep learning approach can be used effectively to aid the Radiologists and Physicians. We have proposed deep learning based automated system which can greatly reduce the time in analyzing HRCT lung images of patients to ensure proper treatment to the needy patients. The system can assists in monitor the progress inpatients recovery by comparing the infection present in the lung.
This paper has been submitted on 14-July-2021. This project work is carried out as part master program in Computer science and Engineering.