| Background Pulmonary cryptococcosis(PC)is a pulmonary fungal disease mainly caused by cryptococcus neococcus and cryptococcus Gattii.It mainly occurs in patients with HIV and low immune function,but also in a small number of patients with normal immune function.Related reports show that the incidence of non-hiv infection of PC increased year by year.The clinical symptoms and signs of PC lack specificity,and there is some overlap in imaging manifestations between PC and lung cancer,tuberculosis(TB)and other pulmonary diseases.On the other hand,cryptococcus has obvious affinity to meninges and nerve tissue,and can cross the blood-brain barrier and cause severe central nervous system infection.If diagnosed clearly at an early stage,the prognosis of patients can be improved and unnecessary imaging follow-up and surgical resection can be avoided.Objective To explore the feasibility of radiomics models for identifying nodule or mass type PC and pulmonary adenocarcinoma,tuberculosis.Methods It was a retrospective analysis of chest computed tomography data of 28 patients with non-HIV nodule or mass type PC,30 patients with pulmonary adenocarcinoma and 26 patients with TB.GE Artificial Intelligence Kitl software was applied to calculate texture feature.ANOVA + MW,Spearman correlation index and Lasso regression were used to filter the main texture parameters,and random forest(RF)machine learning was used for training and testing,and the receiver operating characteristic curve was used to evaluate the diagnostic efficacy both between PC and lung adenocarcinoma group and PC and TB group.In addition,the basic clinical data and CT features of PC,lung adenocarcinoma and TB were observed and analyzed,a semantic feature model and combined mode was established to identify PC from lung adenocarcinoma and TB.Results The age of PC group was lower than that of lung cancer group(P < 0.05).There was no significant difference in the age and sex ratio between the PC and TB groups,neither in sex ratio between PC and pulmonary adenocarcinoma groups.PC and peripheral lung adenocarcinoma groups showed statistically significant differences in morphology,lobulation,burr,pleural depression and satellite focus(P < 0.05).The groups filtered 7 texture features for radiomics modeling,which including the 10 th percentile of histogram,Entropy and Inverse difference moment,Cluster shadow of Gray level co-occurrence matrix(GLCM),and Long Run Emphasis of Run length matrix(RLM).The AUC of the training of the semantic features model,radiomics model,and the combined model of PC and lung adenocarcinoma group were 0.913,0.997,1,specificity of 1,0.964,1,sensitivity of 0.826,1,1.The AUC of the testing was 0.795,0.963,0.950,the specificity of 0.875,0.778,0.9,and the sensitivity of 0.714,1 and 1.There were statistically significant differences between PC and TB group in terms of cavity,halo and pleural depression(P < 0.05).Four parameters such as the 10 th percentile of gray histogram,the correlation of GLCM,and long run emphasis of RLM were selected for radiomics modeling.The AUC of training of PC and TB semantic features model,radiomics model and the combined model were 0.790,0.956,0.944,specificity of 0.938,0.821,0.889,sensitivity of 0.643,0.933,1.The AUC of the testing was 0.583,0.986 and 0.879,the specificity of 0.583,0.889,0.900,and the sensitivity of 0.583,0.875,0.857.Conclusion Radiomics models of CT scan images can be used for identifying nodule or mass PC and lung adenocarcinoma,TB.Radiomics models has higher diagnostic efficiency than CT image,and the combined predictive model of radiomics features and CT semantic features is also has a good diagnostic performance,but not as good as the image of separate radiomics model. |
PDF Full Text Request