Deep Learning Based Pulmonary Nodule Detection Algorithms For Medical Images

Currently,lung cancer has the highest incidence and mortality in cancer diseases.With the change of people's habits and environmental degradation,the population of lung cancer is increasing,thus social concerns are also increasing.Generally,pulmonary nodules are early symptoms of lung cancer in medical images.Although most pulmonary nodules have little effect on human health,a few pulmonary nodules can turn into lung cancer.Once lung cancer emerges,it will cause tremendous threat to human health.Therefore,it is great significant and valuable for prevention and diagnosis of lung cancer to detect pulmonary nodules rapidly and accurately.This thesis uses CT images data as the research object and adopts deep learning methods to study the pulmonary nodule detection algorithm on medical images.The research aims to provide a more accurate and effective diagnostic method for doctors,and improve the survival rate of early lung cancer and create a positive effect on the early diagnosis and treatment of lung cancer.The main contributions of the dissertation are as follows:1.The structural parameters in convolutional neural networks used for feature extraction are redundant.To reduce the degree of the overfitting of network,a new type of network structure of feature extraction on medical images is proposed.By optimizing the network structure parameters,it can extract more effective features.Experimental results show that the proposed new network structure can reduce the overfitting and it is superior to the existing network structure in the detection accuracy of pulmonary nodules.It achieves more accurate and efficient detection of pulmonary nodules.2.It is difficult to detect the shape features of pulmonary nodules in CT images when pulmonary nodules are occluded by blood vessels or weasands.This thesis proposes a novel pulmonary nodule detector by applying the adversarial network to the Faster R-CNN detector.With the characteristics of the new type detector,the adversarial network and the detector can learn from each other.In this way,the performance of detector for pulmonary nodules can be more robust and the total framework realizes the pulmonary nodule detection in an end-to-end network.Experimental results show that the detector based on the adversarial network can improve the detection accuracy of pulmonary nodules.3.For data augmentation method on pulmonary nodule images,this thesis proposes a pulmonary nodule image generation algorithm based on conditional generation adversarial network.This network not only learns a mapping from image to image,but also learns a loss function to train this mapping.Pulmonary nodule medical image data can be generated,which increases the volume of medical image data effectively.Experimental results demonstrate that the detection accuracy is improved when the generated image data is exploited as supplementary training data in training detection model.Thus,the validity of the generated image data is verified.