Research On Classification Of Lung Nodules Based On Deep Belief Network

Due to the living habits of people and the deterioration of environment caused by economic development and other factors,lung cancer has become the highest mortality rate cancer in China and aroused wide attention from all walks of life.With constant improvement of medical level,the most common lung imaging technique,CT technology has become more advanced,leading to lung CT data for each medical personnel has doubled,combined with the number of patients have risen sharply,resulting in huge amounts of CT image.Research indicated that the computer aided diagnosis(CAD)system can provide auxiliary information for the diagnosis of lung cancer,but the final diagnosis depends on the doctor.The explosive increase of data,the diversity of nodules and the working ability of different doctors are the main causes of high misdiagnosis of pulmonary nodules.At present,the CAD system can offer important auxiliary information to doctors.To a certain extent,CAD system can avoid a wide range of diagnostic blindly,and save time and effort for doctors to observe high suspected area.CAD system become an indispensable "assistant" for doctors.Aiming at the complexity of CAD system diagnostic process and the uncertainty of the classification of nodules,the following two aspects are studied in this paper:1.Due to the shape diversity of nodules and the complexity of features,it will lead to over segmentation of nodules in a certain extent.Over segmentation can lead to the loss of effective information,which directly affects the accuracy of the diagnosis.However,if the original CT image(512*512)is used as the input of any learning network,the complexity of the learning process cannot be imagined,and even cannot be completed.In this paper,the target tracking is applied to lung image for the first time.Tracking algorithm,which basing on superpixels is proposed in this paper,is working under the framework of particle filter.It constructs a lung parenchyma appearance model based on superpixel first of all,and then builds the Confidence Map of Images waiting to be tracked and sets the tracking window of the adaptive size.In the process of tracking,real-time updating of templates can ensure the accuracy of them.Reserving the tracking information of every CT in the optimal state among image sequences.Tracking algorithm can quickly pinpoint the area we are interested in,which greatly weaken the interference of superfluous information except for the lung parenchyma in CT images and reduces the complexity of deep learning applying to the diagnosis of lung disease.2.Aiming at the deficiency of traditional classification methods such as BP neural network,support vector machine(SVM)and self-generating neural network(SGNN),we introduce deep belief network into the classification of nodules in this paper.Due to the fact that deep belief network has multiple layers of nonlinear structure,the learning process is composed of supervised learning and unsupervised learning.In this paper,we define a 5 layer depth of deep belief network.Using the lung images obtained in step 1 to train the custom deep belief network.In the training process,the parameters involved in the network are optimized to improve the accuracy of training.