Pulmonary CT Image Analysis Based On Deep Learning

The task of medical image segmentation is to classify medical scanned images at pixels and divide them into disjoint parts to help doctors plan surgical operations.Object detection is to find out the location and size of a specific kind of object in the image,in order to accurately give the region of the object in the image.In the reading of CT medical images,doctors attach great importance to the distribution of pulmonary parenchyma,pulmonary artery,pulmonary airway and other structures.Especially,the images of pulmonary artery,pulmonary airway and other pipelines can provide a lot of anatomical information for doctors.After the location of pulmonary nodules is detected,the spatial structure of trachea,blood vessels and small nodules can be observed in three-dimensional space.In this paper,a medical image processing model for accurate image segmentation and small object detection is proposed,which is based on an automatic image segmentation framework based on deep learning and existing depth neural network structure.The main work of this paper is as follows.(1)Constructed a segmentation network based on the Generative Adversarial Network,proposes an improved training method by using a weighted loss function from cross-entropy and discriminating loss based on gradient penalty,and effectively reduces the number of parameters of the segmentation network by using a variety of dense connection schemes.Comparing experiments on CT sequences from real patients show that the segmentation network shows more than 90%intersection over union scores on the test data from both single block testing and the whole sequences testing after full segmentation experiments.The existing methods need a large amount of parameters and memory for the segmentation network of three-dimensional images.If these methods are applied to the segmentation of three-dimensional visceral objects,the score of the intersection over union can only reach up to 89%,and it is difficult to converge when applied to three-dimensional pipeline segmentation training.In terms of convergence of loss function,the time consuming of converging generative adversarial structure and weighted loss function is 50%of the multi-classification cross-entropy loss training method based on image semantics segmentation,and the training speed is increased by 31.76%.The method used in this paper can be effectively applied in three-dimensional space,and the complex pipeline targets can be segmented.(2)The end-to-end object detection algorithm based on RPN is applied to three-dimensional system.A Dual Path-based algorithm is developed to detect pulmonary nodules.The proportion of DenseNet component is expanded from 1/2 of the total channel to 2/3 of the total channel to enhance the ability of feature exploration.In training,samples are expanded according to three thresholds in each round of training to increase the proportion of nodules and small nodules so that the network can get more training on nodules and small nodules.The experimental results show that the modified structure achieves better performance with fewer parameters like the original DPN.And the FROC score is 1%higher than the ResNet structure used by Faster RCNN,the parameter consumption is only 1/4 of ResNet.In training,fewer parameters can lead to faster training speed,and by expanding the number of samples of nodules and small nodules,more abundant nodule image information can be provided for the detection network.At the same time,the detection network constructed with dual path connection structure is 16%faster than the ResNet structure in each training loop.(3)The implementation details of the algorithm are transparent to doctors and users,and a comprehensive lung CT image analysis system is designed.This paper combines the automatic pipeline image segmentation trained by the improved generative adversarial method with the nodule detection based on the improved DPN network,and carries out detailed system design and module design.Finally,the output image results of each step in the system operation are given,so that the detected nodules can appear in the results together with the segmented pipeline image.