Objective To explore the application value of different deep learning models in predicting the lung cancer spread through air spaces(STAS).Methods A total of 203 patients with stage Ⅰ—Ⅱ primary lung cancer were included,of which 74 were STAS-positive and 129 were STAS-negative.Patients were randomly divided into training set(142 cases)and test set(61 cases)at a 7∶3 ratio.Region of interest(ROI)was outlined using ITK-SNAP software,facilitating the extraction of tumor and peritumoral images.The Resnet18,Inception_v3,and Vision Transformer(Vit)were employed for model training and feature extraction.Feature selection was performed by the least absolute shrinkage and selection operator(LASSO)algorithm and Spearman correlation coefficient,followed by the establishment of a predictive model using the Naive Bayes machine learning algorithm.The receiver operating characteristic(ROC)curve was drawn to compare the prediction performance of each model.The assessment of calibration was performed using calibration curves,and the evaluation of clinical application value was conducted using decision curve analysis(DCA).Results The area under the curve(AUC)for the training and test sets were as follows:the training set Resnet18 0.849-0.930,Inception_v3 0.848-0.888,Vit 0.747-0.842;and the test set Resnet18 0.796-0.846,Inception_v3 0.783-0.804,Vit 0.690-0.796.In tumor-peritumoral images,Resnet18 had a higher calibration and better clinical net benefit,while Vit showed superior calibration and clinical net benefit when only tumor tissue was considered.Conclusion Deep learning models can effectively predict lung cancer STAS,providing more decision support for the preoperative diagnosis and treatment of stages Ⅰ—Ⅱ lung cancer.

Objective To explore the application value of different deep learning models in predicting the lung cancer spread through air spaces(STAS).Methods A total of 203 patients with stage Ⅰ—Ⅱ primary lung cancer were included,of which 74 were STAS-positive and 129 were STAS-negative.Patients were randomly divided into training set(142 cases)and test set(61 cases)at a 7∶3 ratio.Region of interest(ROI)was outlined using ITK-SNAP software,facilitating the extraction of tumor and peritumoral images.The Resnet18,Inception_v3,and Vision Transformer(Vit)were employed for model training and feature extraction.Feature selection was performed by the least absolute shrinkage and selection operator(LASSO)algorithm and Spearman correlation coefficient,followed by the establishment of a predictive model using the Naive Bayes machine learning algorithm.The receiver operating characteristic(ROC)curve was drawn to compare the prediction performance of each model.The assessment of calibration was performed using calibration curves,and the evaluation of clinical application value was conducted using decision curve analysis(DCA).Results The area under the curve(AUC)for the training and test sets were as follows:the training set Resnet18 0.849-0.930,Inception_v3 0.848-0.888,Vit 0.747-0.842;and the test set Resnet18 0.796-0.846,Inception_v3 0.783-0.804,Vit 0.690-0.796.In tumor-peritumoral images,Resnet18 had a higher calibration and better clinical net benefit,while Vit showed superior calibration and clinical net benefit when only tumor tissue was considered.Conclusion Deep learning models can effectively predict lung cancer STAS,providing more decision support for the preoperative diagnosis and treatment of stages Ⅰ—Ⅱ lung cancer.