2142. Looking from the Ground Glass: Can We Discriminate COVID-19 Infections from Other Atypical Pneumonia?
Authors * Denotes Presenting Author
  1. Mutlu Gulbay *; Ankara Sehir Hastanesi
  2. Bokebatur Ahmet Mendi; Ankara Sehir Hastanesi
  3. Orkun Ozbay; Ankara Sehir Hastanesi
The aim of this study is to investigate whether the lesions of COVID-19 and other atypical pneumonia can be differentiated using radiomic parameters.

Materials and Methods:
A total of 134 patients who were diagnosed with COVID-19 (42 male and 22 female; 48,25±13,67 years) or atypical pneumonia (43 male, 27 female; 48,01 ±20,38 years) by clinical and serological tests and had signs of infection on thorax CT were included in the study. Atypical pneumonia group was consisted of influenza A, influenza B, adenovirus, Human coronavirus (HCoV-229E, HCoV-NL63, HCoV-HKU-1), metapneumovirus, respiratory syncytial virus, adenovirus, Mycoplasma pneumoniae, Legionella pneumophila cases. When radiological lesions were grouped as ground glass opacities (GGO), consolidation and crazy paving, 86 patients (37 COVID-19 and 49 atypical pneumonia) had a total of 149 (78 COVID-19, 71 atypical pneumonia) GGO lesions or consolidations with wide GGO halo. All 149 lesions were recorded as Peripheral or Non-peripheral (central and diffuse) according to their localization and segmented volumetrically. Shape (5 parameters), Size (7 parameters), First order (19 parameters) and Second order (14 Gray Level Run Length Matrix, 14 Gray Level Size Zone Matrix, 12 Gray Level Dependence Matrix, 13 Gray Level Co-occurrence Matrix, and 5 Neighboring Gray Tone Difference Matrix parameters) radiomic features were studied in the segmented lesions. All voxels were normalized before calculation and spatial resizing of 0.8x0.8x1.25 mm were performed. Since all studies were performed using 128 slice GE Revolution Evo, no further study harmonization procedure was needed. Textural features extracted using 64 bins.

Although 24 of the 90 radiomic features could discriminate both GGO and consolidations of COVID-19 and atypical pneumonia, they had poor AUC, sensitivity, and specificity as univariate classifiers. The Range parameter had the highest AUC value (0.686), while Lesion Location parameter had the best sensitivity and specificity (60.1% and 68.1%, respectively). To overcome this problem, probabilistic models were created from these 24 parameters. The best 3- and 4-parameter model was determined by comparing Bayesian Information Criterion scores. Models with a higher number of parameters were not created to avoid overfitting. The models obtained were evaluated with logistic regression analysis and leave one out cross validation method. The best model was the combination of Gray Level Co-occurence's Contrast, First Order's Range and Sphericity of Shape features. The AUC value of this model was 0.774, and the specificity, sensitivity and accuracy values in the test set were 76.9%, 78.9% and 77.9%, respectively.

PCR tests can give false-negative results in COVID-19 disease (1). It is recommended to state other viral pathologies may be found in the differential diagnosis even for typical lesions (2). It should be kept in mind that not every pulmonary lesion may be associated with COVID-19 during increased prevalence of other viral infections. Radiomics analysis findings can provide additional data for decision makers.