A Pilot Study Of CT Radiomics And Deep Learning For Predicting Benign And Malignant Pulmonay Solid Nodules

Objective: This study aimed to explore the value of radiomics model and deep learning model based on unenhanced CT in predicting the benign and malignant of pulmonary solid nodules(≤ 2 cm)and further evaluate whether the nomogram can improve diagnostic efficacy by fusing clinical,radiomics,and deep migration learning labeling models.Methods: The clinical data and CT images of 239 surgically treated patients with pulmonary solid nodules(≤ 2 cm)were retrospectively reviewed.Clinical features with p-values less than 0.05 were selected by using t-tests,chi square tests and multiple machine learning models were constructed,finally the clinical labels were established using logistic regression.Features were extracted from CT images,and these features were filtered using normalization,Wilcoxon rank sum test,Spearman correlation coefficient,recursive deletion strategy,Lasso regression.Finally multiple machine learning models were constructed with the final obtained non-zero coefficient features and a logistic regression model was used to establish the radiomics signature.This study integrates and fuses resnet50,resnet101,densenet121,incorporation＿V3,vgg19 and other 5 CNN models,and the final ensemble model was selected as the deep migration learning signature.The fusion of deep learning transformer labels,radiomics labels,clinical labels was performed using logistic regression(LR)to build nomograms and to achieve model visualization.The predictive effectiveness of the above model is evaluated by area under the receiver operating characteristic curve(AUC).Results: Age,gender,smoking history,lobulation sign and spiculation sign were selected as significant features to discriminate between benign and malignant nodules,and the differences were statistically significant(P < 0.05).Clinical labels were established using the above features by logistic regression,and its AUC was 0.753 and 0.750 in the training and test sets,respectively.A total of 1476 features were extracted by pyradiomics software,and 8 radiomics features were finally obtained to establish the radiomics signature with logistic regression,whose AUC was 0.739 and 0.718 for the training set and test set,respectively.Deep migration learning labels were built with the Ensemble model and its AUC was 0.827 and 0.787 for the training and test sets,respectively.The nomogram was built by combining clinical label,radiomics label,deep learning label,and its AUC was 0.880 and 0.843 for the training and test sets,respectively.The AUC in nomogram was significantly higher than that of clinical label,radiomics label and deep migration learning label.However,according to the delong test,the difference in AUC between nomogram and clinical label,radiomics label was statistically significant(P < 0.05),but not between nomogram and deep transformer learning label.Conclusion: In conclusion,the radiomics signature and deep migration learning signature established based on unenhanced CT can effectively predict the benign and malignant of pulmonary solid nodules(≤2 cm).In this study,the combination of radiomics and deep learning by way of nomogram was realized,and the nomogram built by fusing multiple signatures showed the best diagnostic efficacy,which was superior to that of single model,and improved the prediction ability of benign and malignant pulmonary solid nodules(≤ 2cm).