Research On Segmentation Algorithm Of Pulmonary Nodules Based On Multi-View And Semi-Supervised

Precision radiotherapy is one of the mainstream research directions in the combination of medicine and deep learning.In order to segment pulmonary nodules rapidly and accurately,and improve the diagnosis efficiency of doctors,this paper took the pulmonary CT images of LIDC-IDRI dataset and Deep Lesion dataset as the object,studied the improved convolutional neural network segmentation algorithm,and applied a semi-supervised learning strategy for CT image processing.The main research contents and conclusions of this paper are as follows:(1)CT dataset's pre-processing.In order to meet the training needs of the convolutional neural network,all the pulmonary CT images in the public LIDC-IDRI dataset and Deep Lesion dataset were screened first.Subsequently,an improved superpixel segmentation algorithm was used to extract the pulmonary parenchyma in the CT images,removing redundant information such as the torso and bed frame.It would reduce the number of subsequent calculations.Finally,a suitable pulmonary CT dataset for this paper was established through data enhancement methods such as rotation,translation,scaling and flipping.(2)Segmentation algorithm of pulmonary nodules based on multi-scale.For the problems of the high miss rate and rough segmentation edges of pulmonary nodules,a branch of global feature extraction network was designed to provide semantic guidance for low-level features.At the same time,the dense convolution with increased stride size replaced the original maximum pooling layers,enhancing the retention of details;aiming at the large difference in nodule segmentation accuracy of different sizes,the deepest network structure was improved,and three parallel convolutions with different dilation rates were added to improve the generalization ability of the network;for the problem of imbalance in the number of foreground and background samples,the Dice loss function was improved and made the network more inclined to foreground sample data.Experimental results showed that enhancing the extraction of nodule features can effectively improve the recognition ability of the network.Compared with the traditional segmentation algorithms,ours can obtain more fine segmentation edges.Sensitivity of the improved network is increased by 1.56%,achieved a Dice score of 94.37%,and the speed is significantly improved.(3)Algorithm optimization based on multi-view.In this paper,the 3D convolution kernels were used instead of the original two-dimensional convolution kernels.For the problem of large calculation and long time-consuming for pixel-by-pixel nodules segmentation,a detection branch was added before segmentation to reduce the region of interest,optimizing the network structure;To solve the problem that the width-height ratios of the candidate frame didn't match the size of the pulmonary nodules,clustering was performed by the k-means algorithm,and the optimized contour coefficient algorithm was used to determine the corresponding k value.In the CT axial view,pulmonary nodules are easily confused with blood vessels and other tissues.The identification results based on the sagittal and coronal views were incorporated into the detection branch to reduce the number of false-positive samples.Finally,the adjustment factor was incorporated into the cross-entropy loss to improve network accuracy and a regularization term was added to avoid overfitting.It is verified by experiments that the improved neural network in this paper is superior to other three advanced pulmonary nodules segmentation algorithms in terms of the segmentation details and detection accuracy,and it is significantly faster than multi-scale segmentation networks while ensuring the Dice.(4)Optimization of semi-supervised learning strategy.In traditional self-learning,the weights obtained from complex samples are approximate,and the algorithm tends to select nodules with regular shape and single feature.For this problem,this paper optimized the regularization parameter to maximize the most complicated nodule' s weight.In this way,the simpler nodules that are valuable were selected and the possibility of the most complex nodules to be selected was reduced.Aiming at the problem of the low utilization rate of unlabeled samples when only a single strategy was used,an improved semi-supervised learning strategy for pulmonary nodules segmentation network was proposed based on the fusion of self-step and active learning to fully mine unlabeled nodules data.The experiment found that based on iteration,the learning strategy proposed in this paper can achieve high precision with only 500 initial labeled samples.In summary,this paper achieved fast and accurate segmentation of pulmonary nodules based on multi-view convolutional neural network and semi-supervised learning.It has important theoretical and practical significance for promoting intelligent assisted diagnosis in the medical field,reducing the burden on doctors,and achieving "precision radiotherapy" of pulmonary lesions.