Masanori Izumida

Automated seeded region growing algorithm for extraction of cerebral blood vessels from magnetic resonance angiographic data

Magnetic resonance angiography (MRA) has currently played a useful clinical role as a noninvasive method of surveying vascular anatomy. In this study, we developed an automated seeded region growing algorithm for extraction of blood vessels from MRA data. In the conventional region growing algorithm, the user must manually place a seed point within a large blood vessel and also input the segmentation threshold. Furthermore, these processes must be repeated with a different set of seeds and threshold until the satisfactory results are obtained. Thus, this method is time-consuming and the results obtained by this method are highly subjective to the user. With our algorithm, binary images were firstly generated by thresholding the original MRA data to roughly obtain the images of blood vessels, and then the skeletons were generated from these images using the thinning algorithm based on the Euclidean distance transformation. Finally, these skeletons were used as the seeds for region growing. Our method could extract blood vessels automatically and stably, and the segmentation leakage could be largely suppressed. In conclusion, our automated seeded region growing algorithm appears to be useful for extracting and displaying blood vessels in a three-dimensional manner.

Tracking of a Moving Object Using One-Dimensional Optical Flow with a Rotating Observer

The optical flow is a useful tool for the tracking of a moving object. Estimation of the optical flow based on the gradient method is an ill-posed problem. In order to avoid this ill-posed problem, we proposed a tracking method using a one-dimensional optical flow, which is calculated on a straight line (called the calculation axis) spanning several directions. However, the motion of the observer was not considered. In this paper, we propose object tracking by a one-dimensional optical flow under a rotating observer. The apparent motion of a stationary environment object should be eliminated for calculating the one-dimensional optical flow. Hence, we introduce the detection method of a moving object by mapping, which converts the motion of a stationary environment object into a linear signal trajectory. We calculate the one-dimensional optical flow by using pixels, which belong to the moving object, to eliminate the apparent motion of the stationary environment object. In order to verify the efficacy of the proposed method, simulation is performed using synthesized images. The proposed method successfully tracks the moving object when the observer rotates at a constant angular velocity

Estimation of Inverse Kinematics Model by Forward-Propagation Rule with a High-Order Term

Estimation of an inverse kinematics model is important in robot manipulator control. Multilayered neural networks that can express a nonlinear mapping are applied in order to acquire the inverse kinematics model. Several estimation methods using neural networks have been proposed, and we herein consider the forward-propagation (FP) rule. This method is based on the following steps. First, the goal signal, which corresponds to the supervisor signal at the hidden layer and the output layer, is derived by the Newton-like method. Second, the updating of weights is realized by linear multiple regression. The regression coefficients express the adding correction of the weights. In the FP rule, it is important to realize the goal signal exactly. However, it is difficult to realize this requirement because the correction of the goal signal in the hidden layer, which corresponds to the objective variable, is expressed by the linear combination of the desired trajectory, which corresponds to the describing variable, because the neurons in the input layer are decided by the number of control outputs. In this paper, we propose the FP rule with a high-order term. The layer, which is constructed with the high-order term of the input to the neural network, is inserted between the input layer and the hidden layer. In addition, we derive the goal signal in this scheme. In order to verify the efficacy of the FP rule with a high-order term, we apply this rule to the inverse kinematics problem for the two-link manipulator, which moves on the horizontal plane. The simulation results show that the proposed method obtains a more accurate inverse kinematics model, compared to the FP rule without the high-order term

A classification method to reduce the number of categories in ART1

The ART (Adaptive Resonance Theory) neural network was proposed by Grossberg and colleagues to classify by certain criterion (vigilance parameter) a data set distributed across feature space. With ART1, the selection of existing categories and generation of new categories is performed appropriately using two procedures called competition and resonance. Past learning can coexist with new learning without contradictions (that is, the so-called plasticity–stability dilemma is avoided). However, ART classification results are influenced by the data presentation order, so that different numbers of categories and different classification results may be obtained for the same data set, which is basically unacceptable. As a solution to this problem, this study introduces two procedures, namely, asymptotic setting of the vigilance parameter and probabilistic restoration of prototypes. Asymptotic setting of the vigilance parameter is a procedure to reduce the number of generated categories to a minimum by varying the vigilance parameter gradually from its minimum to a preset target value; in the process, the classification results develop from rough to fine. Probabilistic restoration of prototypes handles changes in the order of data presentation by probabilistic variation of existing prototypes (representative values). By adding these two procedures, an ART modification may be designed with a reduced number of generated categories, while not being dependent on data presentation order. In this study, the two additional procedures were introduced into ART1, variety of ART for dealing with binary patterns, and the effectiveness of the proposed algorithm was verified against several data sets. Experiments proved that the proposed method provides order-independent data classification with a minimum (or nearly minimum) number of categories.

Estimation of depth and volume for defects by eddy current testing

The full-length inspection of tubes in steam generator to detect defects using the signal measured by ECT (eddy current testing) is performed in a PWR (pressurized water reactor) plant. According to the technological advances of NDE (non-destructive examination), an estimation of defect shape is required. In this study, by characterization of ECT signals using imitative defects, we express the relation with defect shape in approximate functions, and estimate the defect shape using this function. In our estimation of the depth and volume of defects, the depth estimation error was reduced to 19.7% or less and the volume estimation error to 27.2% or less.

Backpropagation learning algorithm with different learning coefficients for each layer

When considering the algorithm of error backpropagation learning, it is seen that the learning of each layer is not completed independently. This paper proposes a DLBP learning algorithm which uses different learning coefficients for each layer. By changing the ratio of the learning coefficients, the following can be controlled: • the learning progress of the hidden units; • the influence of the hidden units on the network. With DLBP learning, we can construct the corresponding networks to the different requirement such as; • to obtain a network with small number of hidden units; and • to obtain a fault tolerant network.

Analysis of neural network energy functions using standard forms

In this paper, we discuss a method for analyzing the energy function of a Hopfield-type neural network. In order to analyze the energy function which solves the given minimization problem, or simply, the problem, we define the standard form of the energy function. In general, a multidimensional energy function is complex, and it is difficult to investigate the energy functions arising in practice; but when placed in the standard form, it is possible to compare and contrast the forms of the energy functions themselves. Since an orthonormal transformation will not change the form of an energy function, we can stipulate that the standard form represents identically energy functions which have the same form. Further, according to the theory associated with standard forms, it is possible to partition a general energy function according to the eigenvalues of the connection weight matrix; and if we analyze each energy function, we can investigate the properties of the actual energy function. Using this method, we analyze the energy function given by Hopfield for the “travelling salesman problem” and study how the minimization problem is realized in the energy function. Also, we study the mutual effects of a linear combination of energy functions and discuss the results.

Shape Estimation of Defects on Steam Generator Tubes

Most of the nuclear power plants(NPPs) now in service in Japan are based on either of two reactor technologies — PWR and BWR. While the PWR plant requires steam generators(SG), the BWR plant needs no separate SG because of its capability to produce steam directly. On one hand, the need for installation of SG makes the size of the PWR plant larger than that of the BWR plant, while on the other, separate SG installation offers an advantage in that the controlled area in the PWR plant can be limited, resulting in easier plant maintenance, compared with the BWR plant. Accordingly SG maintenance to preclude radioactive substances from flowing out into the secondary system is one of the most important tasks in operating PWR plants.

Selection and consolidation of memorized information for distributed associative memories

Numerous studies have been made on the method of constructing the distributed-type associative memory. This paper discusses the method which selectively memorizes or memorizes by consolidation the information stored in multiple associative memories in a new associative memory. In the distributed associative memory, the information (vector) to be stored is memorized in a distributed way in the form of a matrix (called memorization matrix). Then, in general, it is difficult to know precisely the original memorized information from the already constructed associative memory. This paper proposes a method which can realize the selection and consolidation of the memorized information by constructing directly the memorization matrix of the new associative memory, using the memorization matrices of the already constructed multiple associative memories. The proposed method has the feature that it is not necessary to derive the information memories in each associative memory or to reconstruct the associative memory from the start using the memorized information (relearning). It is a requirement which arises frequently in the construction, update and use of the associative memory to memorize selectively or in a consolidated form the information memorized in multiple associative memories, in an associative memory. For such a purpose, the proposed method will be useful.