Research Of Segmentation And Benign And Malignant Diagnosis Algorithm For Pulmonary Nodules Based On PET/CT

Lung cancer is one of the most serious diseases endangering human life in the world.Clinical studies show that presences of lung cancer is the solitary spherical nodule in lung cavity at early stage.Physicians can make the preliminary analysis of the pulmonary nodules type with the help of medical images.It is hard to determine the benign or malignant pulmonary nodules only by naked eyes because of the different image quality and the lack of the physicians' experiences.It is easy to improve the misdiagnosis and missed diagnosis rate.At present,computer aided diagnosis(CAD)system can be used to analyze the medical images and provide the second opinions for physicians to make the right judgments.CAD can not only reduce the workload of physicians,but also improve the accuracy of the disease diagnosis.Therefore,lung diseases diagnosis with the help of CAD has gradually become a hot topic for many scholars.In CAD system,accurate segmentation is the basis to improve the detection rate of lung cancer,and diagnosis of benign or malignant pulmonary nodules is the necessary way to achieve the rapid recovery.In this paper,the deficiency of pulmonary nodule segmentation and diagnosis methods were analyzed.Based on PET/CT medical images,this paper studies the methods of segmentation and diagnosis of pulmonary nodules:(1)In study of pulmonary nodules segmentation,an algorithm based on LBF active contour model is proposed.This algorithm is mainly aimed at the segmentation of juxta-vascular pulmonary nodules.In CT images,the vessels and pulmonary nodules have high gray similarity,which is prone to edge leakage in process of segmentation.And the circular shape of the vessel transverse may also cause the great disturbance.Moreover,most algorithms have to set the seed points manually andcan not realize automatic segmentation.In proposed method,the region of interest is obtained with the SUV of PET images firstly.The initial contour of pulmonary nodules is constructed by automatic threshold iteration method secondly.After that,energy functional of LBF model is optimized by PET and CT gray level joint vector.It can drive the evolution of initial contour which will stop at the edge of pulmonary nodules accurately.Vessels and nodules can be segmented finnally.Compared with existing algorithms,the proposed algorithm has higher segmentation accuracy and stability,and it can be used as an effective method for juxta-vascular pulmonary nodules segmentation.(2)In study of pulmonary nodules diagnosis,a benign or malignant diagnosis method is proposed based on hybrid restricted boltzmann machine to solve a series of problems in traditional CAD methods,such as features extraction relying on manual design,complex operation and low recognition rate.In proposed method,multilayer unsupervised convolutionl restricted boltzmann machine is applied for features learning from pulmonary nodules images.After that,these features are used as the input of classification restricted boltzmann machine to classify benign or malignant pulmonary nodules.In order to avoid the problem of features homogenization in the process of classification restricted boltzmann machine training,cross entropy sparse penalty is added to optimize it.Experimental results show that this method can avoid the complexity of manual feature extraction effectively and is superior to the traditional diagnosis methods for pulmonary nodules in accuracy,sensitivity,specificity and area under ROC curve values.