Discrimination Of Transient And Persistent Subsolid Pulmonary Nodules Using Deep Transfer Learning

ObjectiveThe purpose of this study was to develop and validate a deep learning model to discriminate transient and persistent subsolid nodules on baseline chest CT,so as to achieve an early diagnosis on the persistence of subsolid nodules.MethodsBy searching chest CT examined from February,2009 to July,2018 in our institution,a cohort of 1414 subsolid nodules from 968 patients,consisting of 319 transient nodules and 1095 persistent nodules were enrolled.The clinical and CT characteristics of each nodules were recorded.The cohort was assigned by a ratio of 70%:15%:15%into a development set(996 nodules,221 transient,775 persistent),a tuning set(212 nodules,47 transient,165 persistent)and a validation set(206 nodules,51 transient,155 persistent).The preprocessing of the images included pixel normalization,patch cropping and data augmentation.Our model was built by transfer learning,which was transferred from a well-performed deep learning model for the malignant/benign classification of pulmonary nodules designed by another research of our department.The performance of the model was evaluated by the AUC,accuracy,sensitivity and specificity of the three datasets,while the performance of the validation served as the true discriminative capacity of the model.The performance of the model was compared to that of two experienced radiologists.Nodules in the validation set were subdivided by Lung-RADS and their size,attenuation type and multiplicity.The performance of each subgroup was calculated to further evaluate the performance and the potential clinical benefit of the model.Occlusion test and t-SNE algorithm were employed to verify the effectiveness of the learned features.ResultsOur model achieved an AUC of 0.926(95%CI:0.889～0.962)on validation set,with an accuracy of 0.859(95%CI:0.804～0.900),a sensitivity of 0.863(95%CI:0.809～0.903)and a specificity of 0.858(95%CI:0.804～0.899),and outperformed the two radiologists.The model performed the best among Lung-RADS 2 nodules and maintained a well performance among Lung-RADS 4 nodules.Feature visualization demonstrated the model’s effectiveness in extracting features from images.In terms of clinical and CT features,there were significant differences on patient sex,nodule size,attenuation type,location and multiplicity between transient and persistent nodules(P<0.05 for all),but no significant difference was found in patient age(P=0.256).ConclusionsThe transfer learning model presented good performance on the discrimination of transient and persistent subsolid nodules.A reliable diagnosis on nodule persistence could be achieved at baseline CT,thus an early diagnosis as well as a better patient care was available.