Application Of Deep Learning In Clinical Diagnosis Of Thymoma

Objective: We intend to find and build a preoperative tool to identify thymoma using enhanced CT with transfer learning,and even further assist thoracic surgeons in determining the nature of anterior mediastinal tumors.Methods: CT images of 167 patients with anterior mediastinal mass were retrospectively selected,and statistical differences were adjusted for age,gender and clinical manifestations in univariate analysis.We trained three different deep learning convolutional neural network models to distinguish thymoma from other benign lesions of the anterior mediastinum and compared their performance at validation.Results: After training and testing with conventional enhanced CT images,the ROC curves of each model after testing were drawn.The AUC of the VGG-19 model for the test set was0.627,of which the 95% confidence interval was 0.516-0.738.The Res Net-18 model for the test set was 0.608 with a 95% confidence interval of 0.494-0.721.The AUC of the Res Net-152 model for the test set was 0.725 with a 95% confidence interval of 0.626-0.823,and there was no statistical difference between Res Net-152 and VGG-19(p=0.196).The VGG-19 model has a specificity of 0.718 and a sensitivity of 0.557;The Res Net-18 model has a specificity of 0.692 and a sensitivity of 0.590;The Res Net-152 model has a specificity of 0.667 and a sensitivity of0.705.The loss of the Res Net-152 model was verified to be 0.5890.Conclusion: In practical clinical applications,it can currently make a preliminary diagnosis,or assist radiologists in diagnosing thymoma with inconspicuous imaging features.When the amount of data is small,over fitting can be reduced to a certain extent by defining regions of interest to strengthen features and introducing the concept of residual network.This study suggests that it is meaningful to further study convolutional neural networks in the diagnosis and treatment of thymoma,and more data are needed to validate our results and train better models.