Classification Of Benign And Malignant Pulmonary Nodules Based On Transfer Learning

Lung cancer is one of the most threatening tumor diseases to mankind.Pulmonary nodules are the early clinical manifestations of lung cancer.Doctors can judge the good and bad attributes of pulmonary nodules through the observation and analysis of lung CT images,and control the patient’s condition in time.Manual observation of CT images to complete the condition judgment is not only a heavy workload and time-consuming,but also the diagnosis results vary from person to person and are subjective.Therefore,introducing deep learning into the classification of benign and malignant pulmonary nodules is of great significance to improve the early screening rate of lung cancer.This paper studies the pulmonary nodule segmentation algorithm based on bbclstmu net network and the benign and malignant classification algorithm of small sample pulmonary nodules based on migration learning,and realizes the accurate classification of benign and malignant pulmonary nodules.The main research contents are as follows:First,the characteristics of the lung nodule image were analyzed,combined with the threshold segmentation method and the morphological method to pre-process the lung nodule dataset,the image was divided into black and white parts through the difference in thresholds,the corrosion operation separated the lung nodule attached to the blood vessel,and the expansion operation filled the small black spot and repaired the small crack,laying a foundation for the subsequent study of the lung nodule segmentation method.Second,aiming at the problem that the Unet model fails to make full use of the weight relationship and semantic information between the coding layers in the lung nodule image segmentation,resulting in insufficient segmentation accuracy,a lung nodule segmentation method based on BBCLSTM-Unet is proposed.This method uses the parallel attention module to complete the larger-scale pulmonary nodule image information encoding and sampling associated channel information weight reset,and uses the BConv LSTM structure to combine feature mapping,fuse forward and backward output sampling to obtain the final segmentation result,and at the same time use the mix loss function to alleviate the class imbalance problem.Experimental results show that the average Dice value of the BBCLSTM-Unet network on the LUNA16 dataset reaches 90.83%,which is better than the original Unet network.Third,aiming at the large difference between the HU value of natural images and CT images of lung nodules,a method of image enhancement of small-sample lung nodules combining autoencoder and K-Means clustering algorithm is proposed.This method introduces an autoencoder to reduce the dimensionality of the natural dataset and the lung nodule dataset,and applies the K-Means clustering algorithm to extract a natural dataset similar to the lung nodule dataset.Using this part of the dataset to complete the pre-training of the VGG16 model,the parameters of the VGG16 model can be closer to the ideal parameters.Fourth,in view of the problem that the number of lung nodule data is small and the deep learning network training needs to use a large amount of data,This paper proposes a classification algorithm for the good and malignant nature of small sample lung nodules based on transfer learning.In order to reduce the risk of overfitting that often occurs in small sample training,this method first performs rotational enhancement of lung nodule images,and improves the VGG16 model to reduce the amount of parameters in the model.Finally,the transfer learning block-by-block fine-tuning of the improved VGG16 model was done using a natural dataset similar to the lung nodule dataset to determine the optimal classification network.Experimental results show that compared with retaining the first,first three,first four and first five convolutional-pooling block parameters,fine-tuning other layer parameters,and retaining the first two convolutional-pooling block parameters,the improved VGG16 network classification effect is the best,which can reach 93.6%.