Lung CAD Algorithm Based On Data Augmentation

Lung diseases,also known as respiratory diseases,include pneumonia,tuberculosis and Coronavirus Disease 2019(COVID-19).According to Forum of International Respiratory Societies,about 334 million people suffer from asthma,and,each year,tuberculosis kills 1.4 million people,1.6 million people die from lung cancer,while pneumonia also kills millions of people.Early detection of lung disease plays a key role in increasing rehabilitation opportunities and improving long-term survival rate,so the correct detection of lung disease is particularly critical.Computed tomography(CT)is widely used to detect and diagnose lung diseases as well as to characterize them and to control their progress,thanks to its high resolution and its low cost.Computer aided detection(CAD)system based on digital image features is essential for comprehensive assessment and early diagnosis of lung diseases.Traditional CAD systems are based on hand-crafted feature extraction engineering such as texture and shape analysis.Hand-crafted features suffer from a deficiency of uniformity,normalization,universality,and require an excessive time-consuming.Convolutional neural networks(CNN)in deep learning has been widely concerned by researchers because of its high accuracy and time-saving.CNN method solves the problems of extraction of relevant features,object recognition,accurate detection and so on.CNN can learn features from data instead of Hand-crafted features based on specific fields.Although CAD based on deep learning has the above advantages,it still has the problems of insufficient training samples and large amount of calculation in the medical field.To solve these two problems,the main contents of this paper are as follows.The data of lung are augmented by using the adversarial and discriminative learning for transferring generative adversarial networks(ADT-GAN).Our idea is to initialize the generator and discriminator of ADT-GAN by using the generative adversarial networks(GAN)pre-trained by dataset A,and then train ADT-GAN on training dataset B.In order to avoid the negative transfer,we add an additional domain discriminator,which gives high scores to the training data and generated data from the target domain,and low scores to the data from the source domain.The purpose is to keep the useful information brought by parameter transferring as much as possible,and make the distribution of the generated image far away from the distribution of the source domain.Our experiments results show that ADT-GAN can yield better training performance,generating better quality samples with fewer iterations.An improved VGG network(T-GAN)is used to detect the presence of pulmonary lesions.We improve VGG16 from network structure and parameter transferring.In the network structure,three full connection layers are simplified to one full connection layer,and batch normalization(BN)is added.At the same time,the parameters of domain discriminator in ADT-GAN are transferred to the network.The experimental results show that compared with VGG16,T-VGG can complete the training in a shorter time,and the accuracy is similar to that of VGG16 in the test set.At the same time,the accuracy of T-VGG on the test set is improved about 4% after using data augmentation.At the same time,the accuracy of T-VGG on the test set is improved about 5% after using data augmentation.