The Key Technologies Research Of Lung Nodule Data Augmentation Based On Adversarial Learning

This research takes the key technologies of lung nodule data augmentation based on adversarial learning as the research object,proposes a lung nodule feature decomposition and shape synthesis algorithm,a lung nodule image reconstruction algorithm and a lung nodule background fusion algorithm,and develops a lung nodule image generation system.This research aims to augment the medical image data by generating realistic and diverse lung nodule images in CT scans.The main contents and achievements of this research are as follows:Firstly,a lung nodule feature decomposition and shape synthesis algorithm based on the adversarial autoencoder structure is proposed.The feature decomposition network can automatically learn how to map the lung nodule image features to the potential representations of attributes and shapes,and the network can learn to obtain feature vectors that conform to the Gaussian distribution through the adversarial learning mechanism.The shape synthesis network can generate lung nodule masks that are similar to the real masks based on the input feature vectors.Secondly,a lung nodule image reconstruction algorithm and a lung nodule background fusion algorithm are proposed.The feature control level of the image reconstruction network is improved by reducing the difference between the predicted feature vector and the input vector through the regression branch,and learning the compatibility between images,shapes and attributes through the feature matching discriminator.The image reconstruction network can generate lung nodule images with corresponding shapes and semantic features according to the input shape masks and attribute feature vectors.The background fusion network can naturally blend the generated lung nodules into the background areas,while preserving the features of the generated lung nodules and the original information of the surrounding background areas.Finally,an end-to-end lung nodule image generation system is designed and implemented based on the proposed algorithms.All networks in the system are optimized by adding the 3D SE-Res Net blocks to improve the feature representation ability and alleviate the difficulty of network training.The system can generate lung nodules of any size with a variety of shapes and attributes at any location in CT scans.The performance of the developed system is verified by building a lung nodule detection network to conduct data augmentation experiments including clinical case detection.The research results show that using the developed lung nodule image generation system to augment the training data of the lung nodule detection network can significantly improve the performance of the network,which finally has highperformance indicators.This research has guiding significance and application value for solving the problem of medical image data scarcity,and has a positive impact on promoting the developments of medical image processing technologies based on deep learning.