RESNET

Nov 28, 2023

Abstract To precisely determine the severity of COVID-19-related pneumonia, computed tomography (CT) is an imaging modality beneficial for patient monitoring and therapy planning. Thus, we aimed to develop a deep learning-based image segmentation model to automatically assess lung lesions related to COVID-19 infection and calculate the total severity score (TSS). The entire dataset consisted of 124 COVID-19 patients acquired from Chulabhorn Hospital, divided into 28 cases without lung lesions and 96 cases with lung lesions categorized severity by radiologists regarding TSS. The model used a 3D-UNet along with DenseNet and ResNet models that had already been trained to separate the lobes of the lungs and figure out the percentage of lung involvement due to COVID-19 infection. It also used the Dice similarity coefficient (DSC) to measure TSS. Our final model, consisting of 3D-UNet integrated with DenseNet169, achieved segmentation of lung lobes and lesions with the Dice similarity coefficients of 91.52% and 76.89%, respectively. The calculated TSS values were similar to those evaluated by radiologists, with an R2 of 0.842. The correlation between the ground-truth TSS and model prediction was greater than that of the radiologist, which was 0.890 and 0.709, respectively. Introduction The rapid pandemic-level outbreak of coronavirus disease 2019 (COVID-19) has caused a wide range and degree of illnesses, predominated by respiratory tract infection 1 , 2 , 3 , 4 . Although most infected patients show asymptomatic or mild clinical manifestations, further investigation beyond real-time reverse transcriptase polymerase chain reaction (RT-PCR) or rapid COVID-19 tests such as chest radiographs is routinely indicated in worsening cases that require hospitalization 5 , 6 . Characteristic findings in chest radiographs of COVID-19 related pneumonia are bilateral patchy and/or confluent and bandlike ground-glass opacity or consolidation in a peripheral and mid-to-lower lung zone distribution. By contrast, several studies have found almost one-half of normal chest radiographs at initial presentation disagree with clinical symptoms 7 , 8 , 9 , 10 . Because of its higher sensitivity, specificity, and speed, chest computed tomography (CT) has become more useful than RT-PCR in early detection, to obtain more information about chest pathology, and to evaluate the severity of lung involvement. Moreover, it can assist triage, especially when hospitalization is required but there is a shortage of healthcare personnel, inpatient beds, and medical equipment, and it may be useful as a standard modality for the rapid diagnosis of COVID-19- related pneumonia 11 , 12 , 13 , 14 , 15 . The chest CT findings are peripheral, bilateral, ground-glass opacity (GGO) with some round shapes with or without consolidation or intralobular lines, a reverse halo sign, or other findings of organizing pneumonia 16 , 17 , 18 , 19 . The total severity score (TSS) has been proposed by Chung et al. 20 . It is calculated from the summation of lesion scores in five lung lobes and is used to categorize the severity of lung involvement and help determine the proper therapeutic management and prognosis 21 , 22 . TSS reflects the clinical classification of COVID-19 22 . It has also been shown to provide high specificity in the detection of severe cases and high inter-observer reliability with a short interpretation time compared to other severity scoring system 23 . It has been used in many studies, such as the comparison of patients with and without vaccination 24 , and the viral load factor for hospitalization and mortality of patients 25 . To reduce the amount of time required for interpretation and increase the accuracy of lesion detection, deep learning has been used to efficiently analyze medical images by performing tasks such as semantic segmentation. Deep learning was also used in the automated assessment of CT severity scores in COVID-19 patients. Lessmann et al. 26 applied deep-learning algorithms that automatically segment the five pulmonary lobes and abnormalities and then predict the severity scores for patients suspected of having COVID-19. The results showed good agreement with the results from independent observers. Chaganti et al. 27 automatically computed the percentage of opacity and lung severity score by applying deep reinforcement learning for lung lobe segmentations and using the U-Net model for a semantic segmentation of GGO and consolidations. The results correlated well with the ground truth. The U-Net model is a convolutional neural network-based model that was originally used for the semantic segmentation of biomedical images and is now one of the most utilized image segmentation techniques. The model structure is U-shaped and consists of two parts: a contracting path (encoder) and an expanding path (decoder) 28 . Subsequently, a U-Net model was created to support three-dimensional (3D) matrices and is called 3D-Unet 29 . The 3D-UNet model was used to develop a more efficient 3D imaging model for the segmentation of lesions and lung tissue 30 , 31 . Cropping the lung area before lesion segmentation can improve accuracy 32 . Enshaei et al. 33 developed a model for predicting the lesion area of COVID-19 patients from CT-scan images, using a model to predict the lung area before the lesion regions were considered. This method enables the lesion model to predict lesions more accurately. In another study, a deep learning model was applied to lung lobe segmentation. The model is capable of accurately segmenting each lung lobe from lung CT scans 34 . It is also utilized in lung lobe segmentation analysis for lung segmentation research to improve segmentation accuracy in multiple diseases such as chronic obstructive pulmonary disease (COPD), lung cancer, and COVID-19-related pneumonia 35 . Many studies have used deep learning models for computer-aided diagnostics to determine the intensity of infections. For instance, Aswathy A. L. and Vinod Chandra S. S. 36 employed 3D-UNet models to effectively segment the lung parenchyma and infected regions in lung CT scans. Additionally, a previous study demonstrated that the effectiveness of these models for medical image segmentation can improve sensitivity performance 37 . In another study, the U-Net model combined with the dense convolutional network (DenseNet) was effectively employed to develop a program for classifying the severity of lung CT in COVID-19 by analyzing the lesion area and comparing it with the lung area in lung CT scan images 38 . They calculated the percentage of infection (PI) using a U-Net model combined with pre-trained models such as residual neural networks (ResNet) and DenseNet. ResNet was first presented by He et al. 39 to solve the vanishing gradient problem of deeper networks by adding feedforward links across some layers, resulting in residual optimization of those layers. DenseNet was first presented by Huang et al. 40 to learn more features by using deeper convolutional layers with many feedforwards linking across layers. For this reason, this knowledge can be applied to lung lobe segmentation and lesion segmentation in CT scan images. In this study, deep learning semantic segmentation was used for the lung severity scoring of the COVID-19 infection. The proposed method utilized a combination of 3D-UNet models integrated with pre-trained models, DenseNet and ResNet, to compute the PI from the lung lobe and lesion segmentation results and estimate the TSS automatically. The aim was to alleviate the radiologist's workload and time spent on imaging diagnostics, as well as improve reporting accuracy. Materials and methods Datasets Due to its retrospective nature, informed consent was waived, and all data were anonymized. This project was approved by the human research ethics committee of the Chulabhorn Research Institute (research project code 167/2564) and complied with the Declaration of Helsinki. These COVID-19 patients were confirmed by RT-PCR acquired from Chulabhorn Hospital who underwent non-contrast enhanced axial chest CT as a part of routine clinical care throughout the pandemic. In this study, we randomly selected 124 cases from the database. The selection contained 28 cases without lung lesions and 96 cases with lung lesions. According to TSS, experienced radiologists classified the cases with lung lesions as mild, moderate, and severe. We divided the selection into 3 groups, i.e., training set, test set 1, and test set 2. The training set was used in model training and validation for lung segmentation and lesion segmentation; test set 1 was for segmentation performance evaluation; and test set 2 was for TSS prediction evaluation. We also randomly selected these cases for each group. In addition, for the training set and test set 1, the numbers of cases across different severity types were set to be equal to prevent class imbalance in the training set (the class imbalance causing a potential bias in the trained model) and for a fair comparison in test set 1. The number of CT slices in these cases ranged from 92 to 208. This information was described in Table 1 . Table 1 Summary of axial lung CT scan datasets.