Feasibility of deep learning technique based on CT radiomics in improving the diagnostic accuracy for pulmonary nodules
10.3969/j.issn.1672-8270.2025.09.003
- VernacularTitle:基于CT影像组学的深度学习技术对提高肺结节诊断准确性的可行性研究
- Author:
Xianhu ZHANG
1
;
Zhigang ZHANG
;
Fang LIU
;
Ying GUO
;
Fan LI
;
Chong LIU
Author Information
1. 保定市第一中心医院影像科 保定 071000
- Publication Type:Journal Article
- Keywords:
Computed tomography radiomics;
Deep learning technique;
Diagnosis for pulmonary nodule;
Accuracy
- From:
China Medical Equipment
2025;22(9):12-16
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the feasibility of deep learning based on computed tomography(CT)radiomics in improving diagnostic accuracy for pulmonary nodules.Methods:A total of 500 patients with pulmonary nodules who admitted to our hospital from January 2023 to January 2024 were selected as study subjects,and they were randomly divided into a training set(350 patients)and a test set(150 patients)as 7:3 ratio.All patients underwent CT examination,and pathological diagnosis was used as gold standard to record pulmonary nodules that were judged by clinical judgment.The radiomics features were screened from the CT images of the patients,and these features were used to construct multiple machine learning models.The predictive value of different models in diagnosing pulmonary nodules was analyzed through confusion matrices and receiver operating characteristic(ROC)curve.Results:A total of 1,594 radiomics features,including 1,195 texture features(74.97%)that was the largest ratio,334 first-order histograms(20.95%),and 65 second-order histograms(4.08%),were extracted in this study.After least absolute shrinkage and selection operator(LASSO)regression analysis and ten-fold cross-validation processing,a total of six radiomics features were screened out.The screened radiomics features were incorporated respectively into four assembled models with machine learning,including ResNet50,DenseNet121,Inception_V3 and VGG19.The constructed models were evaluated respectively using the training set and the test set.The results showed that the assembled model had the highest accuracies in both training set and the test set(96.57%and 95.33%),which area under curve(AUC)values were 0.934 and 0.923,and specificities were 81.64%and 80.52%,and sensitivities were 90.25%and 88.71%,respectively.The results of consistency test indicated that the assembled model had the best classification consistency(Kappa=0.856,P<0.001)in the constructed diagnostic model for pulmonary nodule,which was the best-performing model.Conclusion:The deep learning technique based on CT radiomics has a certain feasibility in improving the diagnostic accuracy for pulmonary nodules,and the machine learning model that is included in this study has favorable predictive value in diagnosing pulmonary nodules.In them,the assembled model that is constructed on the basis of ResNet50,DenseNet121,Inception_V3,and VGG19 has better classification ability.
