Optimization Of PCNN For Medical Image Segmentation

Pulse coupled neural networks (PCNN) is a new type of neural network developed in the1990s, its neuron model is based on the cerebral cortex of cats, monkeys and other animals to the phenomenon of synchronization of impulses. The PCNN has three main advantages that the traditional neural network doesn’t have, including the dynamic of neuron, the synchronization of impulses, and the dynamic of impulses.(1) The dynamic of neuron. Different with the traditional neural network using Weighted input signal and directly compared with the threshold, PCNN is the product of comparing the impulse response function of the input signal with the synaptic channel threshold. The threshold of is neuron not constant, but changes with time, and its change is related with the threshold at previous time and the output of neurons at the current time.(2) The synchronization of impulses. The structure of PCNN is that each neuron has an input (corresponding to a pixel in the image), and the output of neighboring neurons is connected. From the neuron’s point of view, the nerve neuron corresponding to lighter pixels fires faster than the ones corresponding to darker, while on the side of PCNN, at the firing of a neuron, it will be the output sent to the adjacent neurons as their input, which causes neighboring neurons to fire before their natural fire time, and would lead to synchronous oscillations in a large area in the image. Therefore, the behavioral characteristics of PCNN with coupling connection can be described as sync pulses firing with similarity cluster. This means that the neurons with proximity apace and similar brightness to the input fire at the same time, at this time, spatial proximity and brightness characteristics of the clusters are transformed to the image map contains time characteristics. This nature is very important for image segmentation, image automatic target recognition and image fusion.(3) The dynamic of impulses. The characteristic of the dynamic impulses of PCNN neurons is determined by the dynamically variable threshold. The signal generated by multiplying the input signal and the synaptic channel impulse response function is called internal activity, when which exceeds the threshold, the neuron is activated and generates a high output. As the threshold affected by neuron output, this leads the high level of output of the neurons in turn to increase the threshold, so that the internal activity become lower than threshold, and the pulse generator will output zero. Output of this process in neurons clearly form a pulse release (fire), in which the variable threshold characteristics directly affect neuronal inhibition and activation, while the hard limit function determine the output pulse of neurons, the two’s interaction result neuron pulse output. As the pulse frequency and phase are related with neurons’input, so the neuron output can be seen as some sort of frequency modulation or phase modulation of the input signal, which carries the input signal certain features, such features will be very useful for pattern classification and recognition, and image processing.In the last decade, the theory and practice of PCNN have made great progress, replaced the traditional artificial neural networks gradually and become widely used. While conversional PCNN need to set the threshold parameters in the mathematical model, the decay time constant, the weights, and the determination of iteration termination condition is difficult. It’s hard to meet the various needs of image processing, so there is need to optimize the PCNN combining with other related models to achieve better application. Thus, to speed up the PCNN theory and practical application is important and far-reaching significant. The domestic research of PCNN in China started relatively later, most of which is about the theory of PCNN, and the application in image processing, especially image segmentation. Domestic research of PCNN in the field of image processing mainly include:Professor Ma Yide has done a series of studies, and proposed some improved algorithm; Gu Xiaodong, Yu Daoheng also made a thorough research on PCNN applied in image processing, and summarized the dynamic behavior of the PCNN; Professor Zhang Junying also made some research of image processing based on PCNN. In addition, other domestic institutions also done some research of PCNN applied in image processing research. With the development of the research, the degree of attention and study have been deepen. At present, many optimized PCNN models have been proposed.With medical image segmentation as the purpose, a brain MR images segmentation algorithm based on PCNN is proposed, and a lobe segmentation algorithm based optimized PCNN in the chest CT image is designed, the mainly completed work is described as follows:(1) The optimization of traditional PCNNBased on the advantage of the existing optimized methods, a more concise and efficient model is proposed. The main optimization includes the following three aspects:A. Guided by Principle of maximum entropy, the optimized method can avoid the neural network seeking optimal solution through the iterative loop closely which can’t be anticipated.PCNN’s parameters are obtained through repeated adjustments, even in the case of the parameters are selected, the number of loop iterations directly affect the image segmentation result. In order to terminate the algorithm automatically and achieve the best segmentation, principle of maximum entropy is introduced into the method. The greater entropy after segmentation, the more the amount of information retained.B. Optimize PCNN model structure, to make the nerve more sensitive response to external stimuli, while reducing the dependence of PCNN parameters set.The main Optimization includes:(a) The feeding channel only receives external stimuli, to make the physical meaning clearer. And linking channel maintains the natural variation to ensure the accuracy of the algorithm;(b) Since the feeding channel is equal to the external stimuli, the internal activity is improved by plus feeding input and linking input with weight, such change not only keeps the initial state consistent with the original PCNN process, but also maintains the characteristics of PCNN neurons cluster ignition;(c) The pulse generator use constant ignition threshold to reduce the amount of calculation greatly. This threshold is got by using Gaussian distribution histogram in the segmentation process, which has only three peaks of the smooth curve. With the fitted histogram, the best threshold values are got trough the adjacent of two types of issues.C. Adapt weighted neuron connection mode to meet the biological characteristics better.In order to apply structural information the image providing better and improve the efficiency of the algorithm, the linking input is connected with normalized weight model, in which the weight value of the center is slightly bigger than the sum of all adjacent neurons’weights. Thus, central neuron is only affected by the overall impact of adjacent neurons not by individual adjacent one, this makes PCNN a strong anti-jamming.(2) Adaptive3D brain MR image segmentation based on optimized PCNNMagnetic resonance imaging (MRI) with its non-invasive, less affected by movement of object, has been widely used in medical image capture. It’s important to analyze the human brain magnetic resonance images for the study and research of brain anatomy, mental illness. Due to the complexity of the brain MR images and the clinical needs of the application, to segment brain MR images into gray matter(GM), white matter(WM) and cerebrospinal fluid(CSF) precisely and reliably is difficult and hot in modern medical image processing field. Brain MRI without clear boundaries between brain tissues is much complex. As differences between different individuals, as well as the affect of the inhomogenuity of magnetic field in the imaging process, partial volume effect and the effect of noise, it’s complex and difficult to segment the brain. This article firstly use a priori knowledge of the MRI images of the brain tissue to separate the brain with the skull, and then adapt optimized PCNN model to segment the brain tissues into GM, WM and CSF.In image segmentation process, the more the content of the images prior knowledge is used, the more accurate the segmentation algorithm is designed. In general, according to the characteristics of brain MR images, the design of the segmentation algorithm can dependent on the density of the image, tissue location, morphology and other features, and with more priori knowledge used in the algorithm, the probability of image segmentation can be increased.In this paper, the purpose of preprocessing is to separate the brain tissues and the skull and get the best thresholds between GM, WM and CSF. Firstly,3D OSTU is used to remove the background and get the brain and the skull organizing data. Then3D morphological opening operation (corrosion first expansion later) is adapted to remove the connection part between the brain and skull tissues. And then using three-dimensional region growing to get the brain tissue (WM, GM, CSF). Finally, through analysis of histogram of brain tissues data the two threshold value is got.After preprocessing, the complete brain tissue data is got, as well as the two segmentation threshold. Assume brain tissue image as S, the segmentation threshold as T, the maximum entropy value as Hm, the segmentation results corresponding to maximum entropy as Rslt, entropy value for each iteration as Hl, the maximum number of iterations as Nm. The algorithm processes can be described as follows:Setp1:Initialize parameters. Initialize the synaptic weights w; set Fijk[n]=Sijk,Yijk[0]=1,Hm=0, Rsltijk=0. Setp2:Compute the internal activity and output. Lijk[n]=∑WijklmnYlmn[n-1], Uijk[n]=Fijk[n]*Lijk[n], if Uijk[n]>T, Yijk[n]=1, otherwise Yijk[n]=0.Setp3:Computer maximization of mutual information. Get the entropy value H, of current iteration through Y[n], if Hl> Hm, Hm=Hl, and Rslt=Y[n].Setp4:Judge whether the iteration should be terminated. If n> Nm, terminate the iteration, Rslt is the result, otherwise return to step2.Since WM’s density is larger, it’s segmented first and GM is recognized hence from the rest. In addition, the CSF for has no meaning for medical diagnosis, which can be got by removing WM and GM directly from brain tissue data.Experiments show that our three-dimensional PCNN segmentation algorithm is comparatively more concise and more efficient with the traditional PCNN and other brain image segmentation algorithm. The algorithm does not need to set too much the parameters, uses the maximum entropy principle to control the number of iterations, can extract the edge of the image quickly and accurately and retains rich details to guarantee the continuity and integrity of medical image edge with strong adaptability.(3) Lobes segmentation using statistical shape model with PCNN online learningThe lung parenchyma in the chest CT images subdivided into lobe structure is significant for the diagnosis and treatment of lung cancer. But due to partial volume effect (PVE), peripheral space phenomenon (PSP), or the outsize of thickness and so on, the lung fissures may present incomplete or even completely disappear in some layers. And more, the structure of the fissures in the SPECT images is easily mistaken as blood vessels or trachea. So the segmentation result comes out to be unfavorable, which makes it to be a hot research topic in the international arena how to enhance the accuracy and robustness.This article uses statistical shape model (SSM) to segment the lung lobes. The average model of traditional statistical models is generally got via a large number of training samples, which is used in the latter part of the actual segmentation process. This approach requires providing a sufficient number of samples; otherwise the segmentation will be poor. In order to solve this problem, a PCNN-based online training method is proposed which uses the later segmentation result to correct the shape of the model and becomes more representative.The lobe segmentation includes six parts:preprocessing, creation of the initial model, match the statistical model to image, local search of the lung fissures, fine-tuning of the fissures, and the online training of the lung statistical model.To get the shape of the model of the lung parenchyma, automated preprocessing is first applied to the chest CT image, and lung parenchyma contours are obtained with the proposed app-uniform model automatically, obtaining control points of the lung parenchyma contour. At the same time, control points of the fissures in the lung parenchyma are obtained through expert manual segmentation and uniform sampling. Then each sample of lung parenchyma contour control point and lung fissure control points are normalized in the standard coordinate system. Finally, the initial shape of the model is calculated.The actual segmentation process acts as follows:first, the pulmonary is obtained by registration of lung parenchyma contour model and the image. Then the deformation of the registration is applied to the fissure shape of the model, after which, the deformed shape can be used to prognosticate the approximate location of the objective fissure in the image. Second, local research step is executed near the fissure shape of the deformed model, which is designed to find a few pixel on the fissure in the CT images. Last, these pixels are used to make fine adjustments to the shape of the fissure in the deformed model, and get the final lung fissure shape.The segmentation result operated with the same normalized process is pushed into the PCNN online learning system to improve the accuracy of the statistic model continuously. In the PCNN online learning system, the optimized model described as above is used and the different aspects include:(Ⅰ) Each segmentation result is inputted as the external stimulus input.(Ⅱ) The threshold of pulse generator part is dynamic whose initial value is the control points of the statistic shape model, and the output is got by comparing the statistical model and the segmentation result.(Ⅲ) In each iteration, the model shape changes, until the entire network no longer ignite. At this time, the ignition threshold value corresponds to the new statistic shape via the influence of self-learning. With the increase of the segmentation instance, this model will become increasingly accurate and more adaptable.