A Fast Edge Segmentation Of Pulmonary Nodules Based On Deep Learning

Lung cancer is one of the cancers with the highest mortality rate in the world.The main reason is that the average five-year survival rate of patients with lung cancer is less than 5%,and the median survival time of patients with advanced lung cancer is usually 6-8 months.Pulmonary nodules are one of the early manifestations of lung cancer,the risk of worsening disease can be reduced by accurate tracking and prognosis.With the gradual promotion and application of artificial intelligence in the medical field,the labeling and diagnosis of lung CT images can be completed by computer-aided diagnosis systems with high accuracy and efficiency.As the first step of medical image analysis and research,the segmentation of lesion area plays an important role.Therefore,my research focuses on the segmentation of pulmonary nodules.After analyzing the shortcomings of some existing medical image segmentation methods and learning target detection technology shows the advantage of detection speed in the task of the natural image detection,we propose a fast segmentation method based on deep learning,which can improve the segmentation speed and accuracy.We combine the detection technology of lung nodules and the segmentation method of lung nodules into a two-stage segmentation framework of lung nodules.The first stage of the framework is the lung nodule screening module,and the second stage is the lung nodule segmentation module.The second module takes the lung nodule candidate region output in the first stage as input and then segment the candidate region.The two modules train independently without interference with each other.In the lung nodule screening module,in order to increase the detection speed of the lung nodules as much as possible while still having a considerable accuracy,this paper proposes a cross-dimensional lung nodule screening model.First of all,in the two-dimensional rough screening network,we use the 2D fast R-CNN framework,which is integrated into the DUpsampling.By minimizing the loss between the feature map and the compressed tag image,we can get more expressive feature map,and then improve the convergence speed of the network.The high-quality lung nodule candidate area is detected in the entire CT image and its central coordinates are located.The next three-dimensional fine screening network uses the image block as input and uses the 3D spatial pyramid network to screen the detection results of the coarse screening network in the spatial range,so as to reduce the false-positive.The experimental results show that the sensitivity of the lung nodule screening module on LIDC is 90.08%,and the detection time of a single CT image is 340 milliseconds.The pulmonary nodule segmentation module takes the output of the screening module as the input.In order to improve the integrity and precision of the segmentation module for different sizes of pulmonary nodules,this paper proposes a segment based segmentation model of double attention pulmonary nodules.The model integrates two types of attention modules,namely spatial attention module and channel attention module.The spatial attention module makes similar features related to each other by weighting and selectively aggregating the features of each location.The channel attention module selectively emphasizes the dependent channel graph by integrating the correlation features of all channel graphs.Finally,the outputs of the two attention modules are fused to further improve the feature representation.The experimental results show that the model can effectively improve the accuracy of lung nodule segmentation,and the MIo U score on LIDC is 90.44%.In this paper,a two-stage lung nodule edge fast segmentation model is used to segment a single CT image with a lung nodule in 762 milliseconds.To sum up,the lung nodule edge segmentation method based on deep learning proposed in this paper has high performance,which provides a new method and new idea for medical image segmentation.