Automatic segmentation of lungs with pneumonia in X-ray images of patients with COVID-19

Abstract

Some of the features for the novel coronavirus (COVID-19) are severe acute respiratory illnesses with respiratory symptoms including pneumonia. Main motiviation for automatic segmentation and classification of patient images suffering from COVID-19 is reduction of false positive rate to avoid further increasing the burden on the healthcare system. Proper and accurate detection ofdisease would help to provide timely and adequate treatment for the patients affected by COVID-19. Main imaging techniques for pneumonia are CT and X-ray,above which artificial intelligence methods could be added to provide automatic discease detection. Traditional method based on thresholding or active contouring are unable to perform segmentation accurately due to small image contrast and tissue bone overlap (ribs are located over the soft tissue in the lungs area).These limitations,can be overcome usingdeep learning methods such as convolutional neural networks. Up to this date,no high accuracy method been developed yet to detect COVID- 19 pneumonia,we propose U-net to segment the area of lungs. Dataset consisted of 196 radiological images from 25 patients diagnosed with COVID-19.Available dataset was divided into training,validation and testing subsets,for proper learning of the unique features and testing on the unknown images. Results showed that proposed U-net is able to segment lungs well,achieving the dice coefficient of 90.5%. Futureresearch will be directed towards prediction of patient condition in time in order to prescribe adequate treatment timely and in advance.