|Year : 2021 | Volume
| Issue : 1 | Page : 1-8
An artificial intelligence-based radiomics model for differential diagnosis between coronavirus disease 2019 and other viral pneumonias
Mudan Zhang1, Wuchao Li1, Xuntao Yin2, Xianchun Zeng2, Xinfeng Liu2, Xiaochun Zhang3, Qi Chen4, Chencui Huang5, Zhen Zhou5, Rongpin Wang1
1 Guizhou University; Department of Medical Imaging, International Exemplary Cooperation Base of Precision Imaging for Diagnosis and Treatment, NHC Key Laboratory of Pulmonary Immune-Related Diseases, Guizhou Provincial People's Hospital, Guiyang, Guizhou, China
2 Department of Medical Imaging, International Exemplary Cooperation Base of Precision Imaging for Diagnosis and Treatment, NHC Key Laboratory of Pulmonary Immune-Related Diseases, Guizhou Provincial People's Hospital, Guiyang, Guizhou, China
3 Guangzhou Women and Children Medical Center, Guangdong, China
4 Information and Network Management Center, Guizhou Provincial People's Hospital, Guiyang, Guizhou, China
5 AI Lab, Deepwise and League of PhD Technology Co. LTD, Beijing, China
|Date of Submission||25-Jul-2020|
|Date of Acceptance||20-Jan-2021|
|Date of Web Publication||18-Nov-2021|
Dr. Rongpin Wang
Department of Medical Imaging, International Exemplary Cooperation Base of Precision Imaging for Diagnosis and Treatment, NHC Key Laboratory of Pulmonary Immune-related Diseases, Guizhou Provincial People's Hospital, Guiyang, Guizhou
Source of Support: None, Conflict of Interest: None
OBJECTIVE: To set up a differential diagnosis radiomics model to identify coronavirus disease 2019 (COVID-19) and other viral pneumonias based on an artificial intelligence (AI) approach that utilizes computed tomography (CT) images.
MATERIALS AND METHODS: This retrospective multi-center research involved 225 patients with COVID-19 and 265 patients with other viral pneumonias. The least absolute shrinkage and selection operator algorithm was used for the optimized features selection from 1218 radiomics features. Finally, a logistic regression (LR) classifier was applied to construct different diagnosis models. The receiver operating characteristic curve analysis was applied to evaluate the accuracy of different models.
RESULTS: The patients were divided into a training set (313 of 392, 80%), an internal test set (79 of 392, 20%) and an external test set (n = 98). Thirteen features were selected to build the machine learning-based CT radiomics models. LR classifiers performed well in the training set (area under the curve [AUC] = 0.91), internal test set (AUC = 0.94), and external test set (AUC = 0.91). Delong tests suggested there was no significant difference between training and the two test sets (P > 0.05).
CONCLUSION: The use of an AI-based radiomics model enables rapid discrimination of patients with COVID-19 from other viral infections, which can aid better surveillance and control during a pneumonia outbreak.
Keywords: Artificial intelligence, coronavirus disease 2019, radiomics, viral pneumonia
|How to cite this article:|
Zhang M, Li W, Yin X, Zeng X, Liu X, Zhang X, Chen Q, Huang C, Zhou Z, Wang R. An artificial intelligence-based radiomics model for differential diagnosis between coronavirus disease 2019 and other viral pneumonias. Radiol Infect Dis 2021;8:1-8
|How to cite this URL:|
Zhang M, Li W, Yin X, Zeng X, Liu X, Zhang X, Chen Q, Huang C, Zhou Z, Wang R. An artificial intelligence-based radiomics model for differential diagnosis between coronavirus disease 2019 and other viral pneumonias. Radiol Infect Dis [serial online] 2021 [cited 2021 Dec 8];8:1-8. Available from: http://www.ridiseases.org/text.asp?2021/8/1/1/330562
| Introduction|| |
By June 8, 2020, a total of 7,006,436 confirmed coronavirus disease 2019 (COVID-19) patients had been reported globally along with 402,699 deaths (data from coronavirus COVID-19 Global Cases by Johns Hopkins CSSE). The disease continues to present significant challenges as a global public health concern. Being a coronavirus, COVID-19 exhibits similar clinical manifestations to other coronavirus severe acute respiratory syndromes (SARS). Besides coronaviruses, more than 20 kinds of viruses are also the common causes of viral pneumonia infections. Except for medications used in the treatment of influenza-related pneumonia that receive approval from the US Food and Drug Administration, there is limited treatment for noninfluenza respiratory viruses. Establishing early and rapid identification of COVID-19 patients could allow for effective risk assessment and management, while also reducing the economic costs associated with this pandemic.
The Centers for Disease Control and Prevention in the US has recommended to use real-time reverse transcription polymerase chain reaction (RT-PCR) approaches for detecting SARS coronavirus 2 (SARS-CoV-2), and many institutions are improving upon nucleic acid detection technology. However, even rapid antigen detection kits need hours to provide results. Nested PCR and RT-PCR require less time to provide results but are characterized by low sensitivity. Furthermore, PCR is of limited usefulness as the results cannot completely rule out contamination of the specimens. Researchers have recommended in the past that for a precise diagnosis, a PCR examination must be combined with clinical data and radiological characteristics. However, distinguishing between different types of viral pneumonia based on computed tomography (CT) images is not only challenging, but also highly reliant on the expertise of radiologists. Moreover, the appearance of different types of pneumonia in chest CT images is often variable and irregular. Over the last two decades, artificial intelligence (AI) has found wide application in the field of medical imaging. AI methods excel in identifying complicated image patterns and converting them into quantitative data. Researchers have demonstrated that AI models outperformed the majority of radiologists in detecting pulmonary nodules, tuberculosis, and acute conditions pneumonia.
In this study, we seek to establish a radiomics model to identify COVID-19 and contrast the disease's presentation with other viral pneumonias through CT images. A semi-automated lung segmentation technology was applied in the model. The AI model has the potential to be used in large-scale screening programs as well as the individual diagnosis of COVID-19.
| Materials and Methods|| |
Ethical approval and patients
Patients with viral pneumonia are collected from Guizhou Provincial People's Hospital and the radiology quality control center database of Hunan, China. COVID-19 patients come from four medical institutions, namely Jiangjunshan Hospital, Zhongnan Hospital of Wuhan University, Guizhou Provincial People's Hospital and the radiology quality control center database of Hunan province, China. Data distribution of each center is provided [Table 1] in supplementary materials. This multicenter study was approved by the ethics committees of all participating hospitals (2020, NO.01). Because of its retrospective nature, the need to obtain informed consent in advance was waived. The study was performed according to the principles outlined in the declaration of Helsinki.
Inclusion and exclusion criteria
A total of 636 patients were found with 490 of those individuals meeting the inclusion criteria; 392 patients were included in the training and internal test set. The remaining 98 patients came from Hunan's radiology quality control center database and were included in the external test set. The training and internal test sets in this study included hospitalized and nonhospitalized patients with viral pneumonia caused by non-COVID-19 etiologies who had visited the outpatient or emergency departments between January 1 and April 30 of 2019. We selected this timeframe to avoid concurrent overlap with the COVID-19 pandemic. Patients with other viral pneumonias were diagnosed according to confirmed clinical diagnoses documented in daily medical practice (clinical signs and symptoms, viral etiology examination, and CT images reviewed by two radiologists). Patients with incomplete clinical data, low quality or unreadable chest CT images were excluded (n = 17). To avoid influencing our results by detecting the changes in the lungs caused by other diseases, patients were excluded (n = 126) if one of the four following criteria was met: (1) lung tumor (n = 7); (2) autoimmune diseases causing exudation of the lungs including systemic lupus erythematosus (n = 5), rheumatoid arthritis (n = 2), anti-neutrophil cytoplasmic antibodies associated systemic vasculitis (n = 1), lung changes from unknown origin (n = 2), human immunodeficiency virus (n = 3); (3) nonviral pneumonia including fungal pneumonia (n = 2), bacterial pneumonia (n = 17), mycoplasma pneumonia (n = 39), bacterial and mycoplasma pneumonia (n = 4) tuberculosis (n = 13); (4) nonacute inflammatory changes in lung CT images (n = 31).
COVID-19 was diagnosed according to the diagnosis and treatment protocol for the Novel Coronavirus Pneumonia (The General Office of National Health Commission, Office of State TCM Administration, Beijing, China). Patients with incomplete clinical data and low quality or unreadable chest CT images were excluded (n = 3). All 392 patients were divided randomly into training and internal tests at a ratio of 8:2. The external test set in this study included 98 patients, 50 patients with COVID-19 and 48 with other viral pneumonias. [Figure 1] demonstrates the patient inclusion and exclusion criteria.
Chest radiography and clinical data
All images were nonenhanced chest CT images and reconstructed at a slice thickness of 1.00 mm that was retrieved separately from the three hospitals' picture archiving and communication system. Details of the CT characteristics may be found [Table 2] in supplementary materials. Clinical records and laboratory results were obtained from documented medical history.
In the COVID-19 set, CT images were performed within 1 week after patients expressed one or more clinical symptoms such as fever, cough, fatigue, or diarrhea. If the patients had undergone CT scans several times during this period, we selected the first scan. Within 10 days, the final identification of COVID-19 would be confirmed by combining the epidemiologic features (travel or contact history), chest CT, laboratory findings, and RT-PCR for SARS-CoV-2 nucleic acid testing. In the other viral pneumonia set, we chose the first chest CT images performed after hospitalized and nonhospitalized patients visited outpatient or emergency departments.
First, region of interest (ROI) volumes were segmented by an automated segmentation architecture that was based on deep-learning algorithms, which included three modules. Specifically, feature pyramid networks were initially used to obtain the position and area information of the suspicious sign (sign detection). After that, a U-Net network was employed for pixel-level segmentation of the positive regions. Then, the local texture information and shape information of suspicious signs (pneumonia segmentation) were extracted. For the localization and distribution of signs, a three-dimensional (3D)-UNet structure was used to detect sensitivity [Figure 2].
All ROIs were manually modified by a radiologist with more than 5 years of experience and validated by another radiologist with 20 years of experience to further ensure performance quality. The evaluation of the auto segmentation accuracy was finished before image segmentation.
Radiomic feature extraction
To eliminate the intrinsic dependency on voxel size for the radiomic features, a B-spline interpolation resampling was used to normalize the voxel size. Texture features were derived from different directions and scales; thus, the anisotropic voxels scanned at 0.71 mm × 0.71 mm × 0.65 mm or other sizes were resampled to form isotropic voxels of 1.0 mm × 1.0 mm × 1.0 mm. Image normalization was performed using a method that remaps the histogram to fit within μ ± 3 σ (μ: gray level mean between the volumetric interest and σ: gray-level standard deviation [SD]). In addition, for this work we used a fixed bin width of 25 Hounsfield units for discretization in the ROIs.
Radiomic feature extraction was carried out using Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/. Accessed 6 July 2019). Based on the original images, six common feature groups were extracted. The feature groups included first-order features based on the voxel intensity, shape features, and texture features. The latter included the gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM) and gray-level dependence matrix (GLDM). In addition, for pre-processing, we applied filters including wavelet (all possible combinations of applying either a high or a low pass filter in each of the three dimensions, including HHH, HHL, HLH, HLL, LHH, LHL, LLH, and LLL)) and Laplacian of Gaussian (LOG) with different sigma values (1–5 with steps 1). The first-order features and texture features were also extracted from two types of filtered images, including eight wavelet transform and five LoG filters. In total, 1218 radiomic features for each ROI of the pneumonia were extracted from the 3D region of the lesion. After this, the training set was first standardized with the standard scaler package (https://scikit-learn.org/stable/modules/preprocessing.html). Through standardization, the mean of the data was mapped to zero, and the SD was mapped to one. Next, the standardized model in the training set was applied to the test set.
We performed a feature dimension reduction process as high-dimensional features were extracted to select the most relevant features. First, the intra-class correlation coefficient (ICC) was used to assess the reproducibility of the radiomic features measured from the CT images of 30 patients that were derived from randomly proportional sampling from the three hospitals. Next, two independent observers, labeled as Observer 1 and Observer 2, each delineated radiomic features extracted from the ROI twice using the same method. The time interval between the first and second delineation for each observer was approximately 2 weeks. The inter-observer agreement was assessed by comparing the radiomic features extracted from the ROI outlined separately by Observer 1 first, followed by Observer 2. The features showing an ICC > 0.75 indicated satisfactory agreement and were selected for further analysis.
Next, a univariable analysis called K-best was employed. This test selects features according to the K highest scores as computed through the ANOVA F-value between the label and the feature. Features with a significant difference, P < 0.05 were selected. In order to avoid the “curse of dimensionality,” the least absolute shrinkage and selection operator (LASSO) logistic feature-selection algorithm were used to screen the most informative image features.
Model building and statistical analysis
After feature extraction and selection, a logistic regression (LR) classifier was trained to construct a model for distinguishing the two types of viral pneumonia, by using a five-fold cross-validation strategy. The diagnostic performance of the models was evaluated by receiver operating characteristic curves, and the evaluation indicators included area under the curve (AUC), sensitivity (SEN), specificity (SPE), and accuracy (ACC). The optimal cutoff values for any given model were evaluated using the Youden index.
The clinical characteristics were analyzed using the Student's t-test or the Chi-square test with SPSS software (SPSS for Windows, v. 20.0; Chicago, IL, USA). Differences in the AUC values between the models were estimated by the DeLong test. P < 0.05 was considered a statistically significant difference.
| Results|| |
In total, 392 patients were included and randomly divided in the training and internal test set by a ratio of 8: 2 (age, 46.374 ± 25.082 years). Of that total set of patients, 217 (55.4%) comprised the other viral pneumonia set (132 men and 85 women; mean age, 43.180 ± 23.019 years). The other viral pneumonia diagnoses included 51 (23.5%) patients with influenza pneumonia (Type A: 31, Type B: 20); 10 (4.6%) patients with respiratory syncytial viral pneumonia; 4 (1.8%) patients with parainfluenza viral pneumonia; 2 (0.9%) patients with adenoviral pneumonia and 60 (27.7%) patients with Epstein-Barr viral pneumonia. 66 (30.4%) patients suffered from two viral infections, whereas 24 (11.1%) patients suffered from more than two viral infections. The COVID-19 viral pneumonia set contained 175 (44.6%) confirmed patients (86 men and 89 women; mean age, 50.527 ± 18.272). 98 patients from the Radiology Quality Control Center database of Hunan province were included in the external test set that had been previously reported. 50 of them were diagnosed with COVID-19 and 48 patients had influenza pneumonia (Type A: 37, Type B: 11).
Patient characteristics in the training and test sets are listed in [Table 1]. There were no significant differences between the training set and internal test set in age (P = 0.829) and sex (P = 0.912). As such, there is justification for using these two sets as the training set and internal test set.
|Table 1: Clinical characteristics of patients in the training and two test sets (n = 490)|
Click here to view
We performed an analysis of CT scan pneumonia segmentation on 30 patients, who were randomly selected from the entire data set, and the Dice coefficient (DC) was used as the evaluation metric. The DC is defined as:
Where the term ∩ denotes the overlap between the segmented region R1 and the ground truth R2, and |.| denotes the number of voxels belonging to each ROI.
Among the 30 patients, the average DC value was 0.825 ± 0.047, suggesting a solid segmentation result [Figure 3].
|Figure 3: A 53-year-old man with COVID-19. (a) the ground glass nodule exhibited hyper intensity in the chest CT image. (b) the ROI of the nodule was labeled by AI automatic segmentation (yellow color). (c) the ROI of the nodule was corrected manually (red color). The average DC value was 0.82 ± 0.05|
Click here to view
In the intra-reader class, 1131 out of 1218 (93%) radiomic features showed good agreement with ICCs ranging from 0.750 to 0.999, and 87 had ICCs lower than 0.750 ranging from 0.039 to 0.749. In the inter-reader class, 1032 out of 1218 (84%) radiomic features had good agreement with ICCs ranging from 0.750 to 0.997, and 186 had ICCs lower than 0.750, ranging from 0.060 to 0.749. A total of 995 features were selected for further analysis [Figure 4].
|Figure 4: Histogram of the inter-class correlation coefficient and intra-class correlation coefficient. After robustness test, (a) 1032 and (b) 1132 of the initial 1218 CT image features were attained|
Click here to view
Univariable analysis and least absolute shrinkage and selection operator selection of features
In the training cohort, the univariate analysis identified 656 features with statistical significance in association with viral pneumonia. The LASSO LR model was then used to minimize the number of features. The best performance of the LASSO LR was built using a penalty parameter, where λ was 0.0304. With nonzero coefficients, the number of diagnostic features was reduced to 13 [Figure 5]. These 13 features were then evaluated to construct models, including one first-order feature and twelve texture features (GLCM = 7, GLRLM = 1, GLSZM = 1 and GLDM = 3), 10 of the 13 features were transformed by wavelet filters.
|Figure 5: Feature selection using the LASSO binary logistic regression model. (a) By selecting a 10-fold cross-validation in the LASSO model with minimum standards. The binomial deviance was plotted versus log (λ). Dotted vertical lines were drawn at the optimal λ values based on the minimum criteria and 1 standard error of the minimum standards and the optimal λ was 0.0304, with log(λ = -3.4936. (b) The LASSO logistic regression algorithm was used to screen out 13 features with non-zero coefficients out of 656 features. (c) Coefficients in the LASSO model of the 13 features|
Click here to view
Among the 13 features identified, six features were identified as having an associative relationship with COVID-19, and 7 features were closely related to other viral pneumonias. The details of these features are shown in [Table 3] in supplementary materials.
Predictive performance of classifier
We summarized four indicators (AUC, SEN, SPE, and ACC) for AI-based CT radiomics analysis to discriminate COVID-19 from other viral pneumonia in the training and test sets, respectively. In general, LR classifiers achieved a satisfying performance, AUC = 0.91, SEN = 90.2%, SPE = 82.1% and ACC = 86.6% in the training set; AUC = 0.94, SEN = 86.4%, SPE = 88.6% and ACC = 87.3% in the internal test set; AUC = 0.91, SEN = 93.9%, SPE = 74.0% and ACC = 83.8% in the external test set. Delong tests suggested that there was no difference between the training and internal test sets (P = 0.3315); the training set and external test sets (P = 0.3393); the internal test set and external test sets (P = 0.9242). [Figure 6] and [Table 4] in supplementary materials show the results of the LR models.
|Figure 6: ROC curves of LR classifiers. The classifiers achieved AUC of 0.91 in the training set, 0.94 in internal test set and 0.91 in external test set respectively|
Click here to view
| Discussion|| |
Results obtained in this study showed high specificity and sensitivity for the diagnosis of COVID-19 and other viral pneumonias. This noninvasive diagnostic imaging method was efficient and accurate. Furthermore, it had a higher clinical application value, thus making early warning possible in the outbreak of a disease.
In this work, 13 features were evaluated to construct models, 10 of the 13 features were then transformed by wavelet filters. This is a preprint manuscript coauthored by Yang et al. In this article, they proposed a method for the differential diagnosis of COVID-19 and viral pneumonias that were based on radiomics features including GLCM, GLRLM, and GLZLM and had a classification AUC of 0.947. The results showed that the radiomic features were highly distinguishable, which is similar to our conclusion. However, the radiomic features substantially differed from ours, possibly because they did not use filters. Wavelet transformation enabled feature extraction at different resolutions without loss of information. Furthermore, many studies have used features based on the wavelets for classification. Wavelet filters can provide comprehensive spatial and frequency distributions for characterizing ROI in terms of low and high frequency signals. Such features may improve the performance of a radiomics model., In addition, maximum (original-first order), the maximum gray level intensity within the ROI, and GLCM_IMC_1 (original) were also extracted in our study. The maximum has more to do with the severity and consolidation. This likely is the simplest feature someone without sufficient experience may detect with the naked eye. IMC represents texture complexity, where the greater the value equates with increased uniformity or homogeneity. The coefficient of IMC was −0.0088, indicating other viral pneumonia with higher homogeneity and internal heterogeneity. Because the analysis of textures is subjective and unsupported by evidence, we can only glean some clues from pathology and imaging.
The pathological features of COVID-19 greatly resemble those seen in SARS and Middle Eastern respiratory syndrome coronavirus infections based on what histological examinations have shown. However, systemic anatomical observation has revealed that pulmonary fibrosis and consolidation in COVID-19 are not as severe as what was seen in SARS, despite COVID-19 causing a more obvious exudative response than SARS. Moreover, the first COVID-19 autopsy case reported in China revealed that a large amount of white and sticky secretions spilled from the alveoli of inflammatory lesion areas that corresponded to the ground-glass lesion seen on CT images; the entire lung lost its sponginess. In past research, CT findings of viral pneumonia have been diverse and may be affected by the immune status of the host and the underlying pathophysiology of the viral pathogen. Our research indicates that COVID-19 has its own pattern of texture characteristics seen in microscopic image textures and may be characterized by greater heterogeneity but further proof and verification are needed.
| Conclusions|| |
To summarize, in this study, we analyzed CT imaging data using a machine learning approach based on radiomic features. The software we used in our study is advantageous in terms of it applicability and availability. However, it must be noted that this study had several limitations. First, the applicability of our findings may vary depending upon the relative prevalence of the etiological agent in any particular period. Second, cases of patients with COVID-19 and other concurrent viral pneumonia infections may not have been easily discoverable during the COVID-19 outbreak. Finally, the radiomic features may have been influenced using different scanners, radiological parameters, slice thicknesses, and equipment in different hospitals.
We would like to thank LetPub (www.letpub.com) for its linguistic assistance during the preparation of this manuscript for free during COVID-19 pandemic.
Financial support and sponsorship
This study was supported by the Guizhou Science and Technology Project (QKHZC(2020)4Y002, QKHPTRC(2019)5803), the Guiyang Science and Technology Project (ZKXM(2020)4), Guizhou Science and Technology Department Key Lab. Project (QKF(2017)25) and National Natural Science Foundation of China (81960314).
Conflicts of interest
There are no conflicts of interest.
| Supplementary Materials|| |
| References|| |
Luo G, Gao S. Global health concerns stirred by emerging viral infections. J Med Virol 2020;92:399-400.
Dandachi D, Rodriguez-Barradas M. Viral pneumonia: Etiologies and treatment. J Investig Med 2018;66:957-65.
Udugama B, Kadhiresan P, Kozlowski H, Malekjahani A, Osborne M, Li V, et al.
Diagnosing COVID-19: The disease and tools for detection. ACS Nano 2020;14:3822-35.
Ai T, Yang Z, Hou H, Zhan C, Chen C, Lv W, et al.
Correlation of chest CT and RT-PCR testing for coronavirus disease 2019 (COVID-19) in China: A report of 1014 cases. Radiology 2020;296:E32-40.
Cajal S, Hümmer S, Peg V, Guiu X, De Torres I, Castellvi J, et al.
Integrating clinical, molecular, proteomic and histopathological data within the tissue context: Tissunomics. Histopathology 2019;75:4-19.
Shi H, Han X, Jiang N, Cao Y, Alwalid O, Gu J, et al.
Radiological findings from 81 patients with COVID-19 pneumonia in Wuhan, China: A descriptive study. Lancet Infect Dis 2020;20:425-34.
Koo H, Lim S, Choe J, Choi S, Sung H, Do K. Radiographic and CT features of viral pneumonia. Radiographics 2018;38:719-39.
Chassagnon G, Vakalopoulou M, Paragios N, Revel M. Artificial intelligence applications for thoracic imaging. Eur J Radiol 2020;123:108774.
Ga¨el Varoquaux AG. Scikit-learn: Machine learning in python. J Mach Learn Res 2011;12:2825-30.
Zhao W, Zhong Z, Xie X, Yu Q, Liu J. CT scans of patients with 2019 novel coronavirus (COVID-19) pneumonia. Theranostics 2020;10:4606-13.
Yang N, Liu F, Li C, Xiao W, Xie S, Yuan S, et al. “Diagnostic classification of coronavirus disease 2019 (COVID-19) and other pneumonias using radiomics features in CT chest images”, Artif. Intell. Mach. Learn 2019:1-11.
Chun S, Suh Y, Han K, Park S, Shim C, Hong G, et al.
Differentiation of left atrial appendage thrombus from circulatory stasis using cardiac CT radiomics in patients with valvular heart disease. Eur Radiol 2021;31:1130-9.
Chaddad A, Daniel P, Niazi T. Radiomics evaluation of histological heterogeneity using multiscale textures derived from 3D wavelet transformation of multispectral images. Front Oncol 2018;8:96.
Majeed A, Hameed A, Mohd H, Hamed A. Utilizing hybrid functional fuzzy wavelet neural networks with a teaching learning-based optimization algorithm for medical disease diagnosis. Comput Biol Med 2019;112:103348.
Das D, Dutta P. Efficient automated detection of mitotic cells from breast histological images using deep convolution neutral network with wavelet decomposed patches. Comput Biol Med 2019;104:29-42.
Xu Z, Shi L, Wang Y, Zhang J, Huang L, Zhang C, et al.
Pathological findings of COVID-19 associated with acute respiratory distress syndrome. Lancet Respir Med 2020;8:420-2.
Ruuskanen O, Lahti E, Jennings LC, Murdoch DR. Viral pneumonia. Lancet 2011;377:1264-75.
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5], [Figure 6]