Research On CT Diagnosis Algorithm Of Solitary Solid Pulmonary Nodules Based On Generative Adversarial Network

As one kind of cancers with the highest morbidity,lung cancer seriously endangers human life safety and health.Early detection and diagnosis of lung cancer helps provide the best rehabilitation and treatment plan for patients.However,lung tuberculosis and lung adenocarcinoma can both clinically appear as solid pulmonary nodules,leading to misdiagnosis and missed diagnosis,which brings great challenge to preoperative diagnosis of solid pulmonary nodules.As deep learning rapidly develops,existing studies have shown that deep learning technology has achieved good results when applied to medical image processing.Nevertheless,in the field of computer aided diagnosis based on deep learning,clinically obtained dataset of solid pulmonary nodules suffers from data scarcity,causing deep learning models to overfit the dataset.Meanwhile,labelling clinical dataset manually and precisely requires high time and economic costs,bringing disadvantages to the training procedure of supervised learning models.Aiming at solving the problems mentioned above,in this dissertation we proposed a feature extraction algorithm and a diagnosis algorithm for CT images of solid pulmonary nodules based on generative adversarial network to classify lung tuberculosis and lung adenocarcinoma that both behave as solid pulmonary nodules.Firstly,to solve the problem of data scarcity,a diagnosis algorithm for CT images of solid pulmonary nodules based on generative adversarial network and attention mechanism was proposed.This algorithm augmented clinical dataset using generative adversarial network,and then combined convolutional neural network with attention mechanism to build a feature extraction model based on Dense Net and Transformers structure.The feature extractor was trained on the augmented dataset to extract deep features of solid pulmonary nodules.To further improve the accuracy of pulmonary nodules malignancy diagnosis,the proposed algorithm combined deep features with subjective features.Results on clinical datasets indicated that the proposed diagnosis algorithm relieved the degree of overfitting of end-to-end model on small datasets and helped improve the classification performance between lung tuberculosis and lung adenocarcinoma.Secondly,to solve the problem of lacking manually labelled data,a feature extraction algorithm for CT images of solid pulmonary nodules based on generative adversarial network and self-supervised learning was proposed.The network structure of the proposed algorithm was constructed based on generative adversarial network while the idea of multi-task learning was introduced into the algorithm.The discriminator in generative adversarial network ought to complete semi-supervised classification tasks and self-supervised pretext tasks along with the original adversarial task.The well-trained discriminator would be used as a feature extractor in downstream diagnosis task.Results on clinical datasets indicated that the proposed feature extraction algorithm could provide effective features of CT images of solid pulmonary nodules using only a small amount of manually labelled data.Features extracted by the proposed algorithm had good stability and robustness,which relieved the impact of data scarcity on discriminator and could achieve better performance when features were applied to downstream diagnosis task.