خلاصة:
The continuing outbreak of COVID-19 pneumonia is globally concerning. Timely detection of infection ensures prompt quarantine of patient which is crucial for preventing the rapid spread of this contagious disease and also supports the patient with necessary medication. Due to the high infection rate of COVID-19, our health management system needs an automatic diagnosis tool that equips the health workers to pay immediate attention to the needy person. Chest CT is an essential imaging technique for diagnosis and staging of 2019 novel coronavirus disease (COVID-19). The identification of COVID-19 CT findings assists health workers on further clinical evaluation, especially when the findings on CT scans are trivial, the person may be recommended for Reverse-transcription polymerase chain reaction (RT-PCR) tests. Literature reported that the ground-glass opacity (GGO) with or without consolidation are dominant CT findings in COVID-19 patients. In this paper, the U-Net based segmentation approach is proposed to automatically segment and analyze the GGO and consolidation findings in the chest CT scan. The performance of this system is evaluated by comparing the auto-segmented infection regions with the manually-outlines ones on 100 axial chests CT scans of around 40 COVID-19 patients from SIRM dataset. The proposed U-Net with pre-process approach yields specificity of 0.91 ± 0.09 and sensitivity of 0.87 ± 0.07 on segmenting GGO region and specificity of 0.81 ± 0.13 and sensitivity of 0.44 ± 0.17 on segmenting consolidation region. Also the experimental results confirmed that the automatic detection method identifies the CT finding with a precise opacification percentage from the chest CT image.
ملخص الجهاز:
The performance of this system is evaluated by comparing the auto-segmented infection regions with the manually-outlines ones on 100 axial chests CT scans of around 40 COVID-19 patients from SIRM dataset.
The popular DL based segmentation models such as a fully convolutional neural network (FCN) (Shelhamer, Long, & Darrell, 2017), Deep encoder and decoder model for image segmentation (Badrinarayanan, Kendall, & Cipolla, 2017), and U-Net convolutional network for medical image segmentation (Ronneberger, Fischer, & Brox, 2015) have been utilised in various medical image segmentation applications.
al (Agnes, Anitha, & Peter, 2018) have proposed a U-Net based Convolutional Deep and Wide network (CDWN) model to separate the lungs from chest CT scans accurately without any post-processing operations.
HU value histogram of 50 arbitrary axial slices of CT scans of COVID-19 cases from the SIRM training set: HU distribution of the lungs and infection in the whole CT image (left) and HU distribution of GGO and consolidation in the infected region (right).
For this experiment, the lung regions are segmented from the input chest CT scan images using the existing convolutional deep and wide network (CDWN) model (Agnes et al.
Evaluation results of U-Net with and without pre-processing on SIRM Dataset {مراجعه شود به فایل جدول الحاقی} (View the image of this page) SIRM Case 5 - Slice position:88, Patient Age:64, Gender:Male Manual Diagnosis - Opacification percentage of the lung: 88.
Automatic Chest CT Image Findings of Novel Coronavirus Pneumonia (COVID-19) Using U-Net Based Convolutional Neural Network.
Automatic Chest CT Image Findings of Novel Coronavirus Pneumonia (COVID-19) Using U-Net Based Convolutional Neural Network.