Research On Benign And Malignant Classification Of Lung Nodules By CT Based On Deep Learning Radiomics

The early classification of benign and malignant pulmonary nodules,which is one of the malignant tumors of pulmonary nodules,is crucial for the follow-up treatment.The DLR model classifies by combining manual features,radiological features and depth features,making full use of features of multimodality and contributing to enhancing the diversity of features.Nevertheless,DLR prediction still has many problems,such as low segmentation accuracy,small CT data volume,small target area.This thesis mainly focuses on the problems of low segmentation accuracy,small data volume and low classification accuracy.The main work is described as follows:First,point at the problem of low segmentation accuracy,We propose an improved UNet network for segmentation.The network takes continuous CT slices as the input to capture the characteristic relationship between continuous slices centered on nodules.Similar to that,physicians will refer to adjacent slices when sketching ROI regions,thus improving the segmentation accuracy.The improved network is verified on the LUNA16 public data set.The experiment shows that the Dice and MIo U of the improved model can reach 85.56% and84.72% respectively,which is 5.97% and 5.91% higher than that of UNet.Secondly,an improved DLR method for CT classification of benign and malignant pulmonary nodules is proposed.In the part of depth feature extraction,aiming at the small amount of CT data,a method of depth feature extraction based on transfer learning is proposed.This method solves the problem that the neural network needs a lot of data training through pre-training and fine-tuning strategies.In the part of modeling and analysis,aiming at the problems of the stability and generalization ability of a single model,a modeling method of model integration is proposed.This method integrates multiple predictors,which can learn from each other and improve the stability and accuracy of the results.The improved model was verified on the LIDC-IDRI public data set.The experiment shows that the classification index AUC and ACC reach 94.2% and 92.5% respectively,which were improved by more than 2% compared to other methods and have excellent ability to distinguish benign and malignant nodules.