Analysisre And Engineering Realization Of Lung CT Image Processing Basedon Deep Learning

Lung Lung cancer is one of the most deadly cancers in the world.It is very important to use image processing technology to study the diagnosis of lung cancer.The traditional method is based on the lung CT image segmentation,artificial naked eye lung CT diagnosis and detection.In recent years,image processing technology has been widely used in lung cancer recognition,especially the continuous development of deep learning.Using deep learning to automatically recognize lung cancer will save a lot of time for radiologists and make diagnosis more economical.Generally,the general method to detect lung cancer is to segment the lung first,then detect and recognize the features.In this thesis,aiming at some difficulties in lung CT image segmentation and recognition,such as the complexity and variability of lung node region structure,combined with the recent deep learning technology,the feasibility of using the related technology of Unet segmentation and 3DCNN recognition is studied and realized in engineering.The main work of this paper includes:(1)In this thesis,Compared with other deep learning models,the Unet model is simple in structure,small in parameters,fast in segmentation and has very good generalization performance,and has high segmentation accuracy for many different data sets.In this paper,based on the above research,the lung node segmentation based on Unet is realized in engineering.It is mainly composed of four modules:data set reading module,Unet segmentation algorithm module,interactive event module and result display module.Firstly,clinical scanning lung CT is used The results show that the deep learning algorithm is better than the machine learning algorithm in both ACC and IOU.When acquiring the data set,the gray level is blurred,the noise redundancy is high,and the pixel range is wide because of the location and edge of the machine,In order to enhance the clarity and contrast of the image,the filter and data correction are added to preprocess the data.(2)In this thesis,3DCNN network structure is studied.3D convolution is more efficient than 2D convolution in automatic extraction of 3D image feature information,which better represents the spatial and texture characteristics of 3D image.In this paper,based on the above research,the lung cancer detection of 3DCNN is realized,which is mainly composed of four modules:data set reading module,recognition algorithm module,interactive event module,and result display module The results show that 3dcnn achieves good results in the process of three-dimensional image classification of pulmonary nodules.Good specificity means that more malignant pulmonary nodules can be detected in the same data set.The introduction of FROC to this paper’s 3DCNN model is used to evaluate the false-positive pulmonary nodules in the data set,combined with morphology and pattern recognition to detect lung cancer.