畑 豊

© 2002 Nanyang Technological University. Augmented reality (AR) is a combination of a real scene viewed and a virtual scene generated by a computer. This paper introduces fuzzy inference techniques into AR. The fuzzy inference based AR enhances a region of interest (ROI) in medical images using expert knowledge expressed with fuzzy if-then rules. Therefore, the AR enables users who are not familiar with such medical images to observe the ROIs as well as expert's observation. Especially this paper discusses the fuzzy inference based AR in MR Cholangiopancreatography images. The AR enhances tube-formed tissues such as the pancreatic duct. In this case, a step for finding the tissue using fuzzy inference is included. We apply the proposed the fuzzy inference based AR to three subjects. Through experimental results, we showed the pancreatic duct was successfully found from any subjects and could clearly augment the 3D shape of the pancreatic duct.

© Springer-Verlag Berlin Heidelberg 2002. In medical images, the tube-formed tissues such as the blood vessels, the trachea, and the pancreatic duct are sometimes partially masked because of the constriction, stones in the vessels, the pancreatic cancer, etc. Therefore, it is not easy to automatically segment the region of tubes (ROTs) from medical images for visualizing the structures by using conventional image segmentation methods, because inference of ROTs is difficult. In this paper, we propose a fuzzy rule-based augmented reality method for finding non-continuous ROTs. We can obtain the ROT without extracting it. The physicians’ procedure for finding the ROT can be eliminated by fuzzy inference techniques based on their knowledge. The employed knowledge is the intensity, the curve, and the radius of the ROTs. We apply the proposed method for finding the pancreatic duct from MR Cholangiography images. Through experimental results, we show that this method can successfully find the pancreatic duct from any data sets and it can clearly visualize the 3D shape of the ROT in MIP images.

Near-infrared spectroscopy (NIRS) is a recently developed method, which can investigate the human brain function with noninvasive, high time resolution, and high portability. It has been applied to assessment of language dominance by Watanabe et al, however, it has a risk which some measurement points took other brain activities occurred in like the motor area. This paper improves the method by introducing fuzzy logic. Fuzzy logic is used to giving fuzzy degrees belonging to the speech, motor and other areas for each measurement point. Using the obtained fuzzy degrees, we redefine the laterality index as the fuzzy mean of the laterality indexes for all paired measurement points.

This paper describes comparison between our developed two Non-Destructive Testing (NDT) system with ultrasonic method and pulse-radar method. These systems could provide information of embedded objects in concrete structure. First, we introduce these NDT systems. The raw output images of these methods keep us from understanding them. To overcome this problem, we apply fuzzy logic into them. Next, we compare these two developed systems to show how to select appropriate method for various cases. Consequently, pulse-radar system is good for general purpose, because this system needs no coupling medium. Ultrasonic system is appropriate for the purpose (e.g. to detect both embedded objects and cracks, to examine deeper area, and in case of gathering of reinforcing bars). We recommend that these two systems are complementally for realizing NDT system.

This paper proposes an application of a Genetic Algorithms (GA) to Fuzzy Non-Destructive Testing (NDT) System. We are concerned here with the system extracts embedded tubes from pulse-radar images. The system usually uses several fixed parameters to analyze input data. The optimal values of the parameters depend on the environment (e.g. permittivity of concrete, pulse-radar device, and so on). Therefore, the system often fails extract embedded tubes when the fixed parameters are not optimal in the environment. No work deals with this problem. This paper attempts to optimize them by a simple GA in order to solve the problem. We applied the system with the GA to two data sets obtained by various environments. As the result, the system with the GA was able to adapt each environment.

The information obtained from lacustrine laminated diatomite gives us the environmental records in the past. The varved diatomite is composed of the alternation of light-colored and dark-colored laminae. The lamina thickness has the periodic variation, which is most likely affected by climatic change due to periodic solar activity. In this paper, we propose an effective method for evaluating the space-varying lamina thickness of lacustrine diatomite. The proposed method is achieved by using 2-dimensional(2D) filters and space/spatial-frequency analysis. Through the simulation of lacustrine laminated diatomite, we show that this simple imaging algorithm allows clear and intuitive display of the change in the lamina thickness with respect to space (the distance from the surface).

In this paper, we describe a development of a new ultrasonic oscillosensor and its application to recognition of human action. The ultrasonic wave is used in many fields due to its low cost and non-invasiveness. The ultrasonic wave sensitivity catches the difference of the sound property and it propagates in the water. By using this, we develop a new oscillosensor using the ultrasonic wave. Firstly, this paper shows the principle of this new sensor. Secondly, as an example of the practical application, this paper describes a recognition system of human action on the medical bed. From the result, this sensor can monitor the vital action of a patient and it can detect a state that a patient falls into bad condition, such as motionless or suffering.

In transcranial sonography system, we usually placed the sensor to anterior and superior of the attachment of the upper ear (posterior temporal window). Due to this limitation, we cannot obtain transcranial information from arbitrary place of the skull. Although this limitation free system is strongly required in clinical treatment, there is no discussion of the placement free transcranial sonography system. This paper describes a new transcranial system without this limitation of the placement. We did the experiment for investigating to identify the tissue under the artificial bone, which assumes the skull and cerebral hemisphere. In it, we employ a fuzzy inference technique to achieve high identification robustness. This system is available for radiation free bedside technique, emergency usage, and non-invasive usage for diagnosing brain diseases.

© 2002 Nanyang Technological University. We propose a new algorithm called the representative line detection algorithm which embeds physician knowledge with fuzzy if-then rules. The algorithm detects the representative line of the region of interest (ROI). The representative line can show the rough location and shape. We first consider the representative line which consisted of some nodes. These nodes are then automatically detected by tracking the most suitable direction from the starting node. To evaluate this algorithm, it is applied to segmentation of the meniscus from CT knee images. The experimental results of six normal subjects showed that the representative line detection algorithm could express the center line of the meniscus, and could lead to detection of successful segmentation of the menisci.

Augmented Reality (AR) is a combination of a real scene viewed and a virtual scene generated by a computer. This paper introduces fuzzy inference techniques into AR. The fuzzy inference based AR enhances a region of interest (ROI) in medical images using expert knowledge expressed with fuzzy if-then rules. Therefore, the AR enables users who are not familiar with such medical images to observe the ROIs as well as expert's observation. Especially this paper discusses the fuzzy inference based AR in MR Cholangiopancreatography images. The AR enhances tube-formed tissues such as the pancreatic duct. In this case, a step for finding the tissue using fuzzy inference is included. We apply the proposed the fuzzy inference based AR to three subjects. Through experimental results, we showed the pancreatic duct was successfully found from any subjects and could clearly augment the 3D shape of the pancreatic duct.

This study was conducted to determine how mobile phones (MP) and the Internet through personal computers (IPC) have been used among Japanese elderly women. It also aims to examine the relationship between the input devices and factors such as age, health status, and future needs for a MP and the IPC. The magnitude of the relation was evaluated by the odds ratio (OR). The survey was done in December 2000 and January 2001. The study sample was forty women, 78.5 +/- 4.5 years of age (70-89), who lived in their own houses in Y Town near Himeji City, Hyogo, Japan. Findings reveal that use of a MP was 10.0% and none used the IPC. On the other hand, 60.0% of the subjects were interested in using a MP in the future and 27.5% in using the IPC. The touch screen was the most preferred input device among the elderly with high age [OR = 3.86, 95% confidence interval (CI) 0.83-18.98]. It was also preferred by subjects with more difficulties (OR = 5.00, 95% CI 1.05-25.41) and less future need for a MP and the IPC (OR = 7.22, 95% CI 1.34-43.88). On the other hand, the Japanese kana syllabary (JKS) was the most preferred input device among those with a low age (OR = 4.33, 95% CI 0.68-35.00) and with a more future need for a MP and the IPC (OR = 6.33, 95% CI 1.14-39.59). Considering these results, we intend to create a type of keyboard that combines the features of touch screen and JKS for the elderly women.

In this paper we propose a learning algorithm to enhance the fault tolerance of feedforward neural networks (NNs for short) by manipulating the gradient of sigmoid activation function of the neuron. We assume stuck-at-0 and stuck-at-1 faults of the connection link. For the output layer, we employ the function with the relatively gentle gradient to enhance its fault tolerance. For enhancing the fault tolerance of hidden layer, we steepen the gradient of function after convergence. The experimental results for a character recognition problem show that our NN is superior in fault tolerance, learning cycles and learning time to other NNs trained with the algorithms employing fault injection, forcible weight limit and the calculation of relevance of each weight to the output error. Besides the gradient manipulation incorporated in our algorithm never spoils the generalization ability.

This paper proposes a fuzzy expert system scheme for foot CT image segmentation. Our method aims at decomposing a fractured foot bone into each of fragments and normal bones, as a framework of developing a computer-aided orthopedic surgery planning system. It employs fuzzy knowledge representation to express orthopedic surgeon's knowledge: knowledge of distance to extract fragments and knowledge of intensity distribution to decompose adjoining normal bones on the articular cartilage. The two types of knowledge are compiled to a total degree, which is evaluated in the segmentation procedure. Qualitative evaluation of an orthopedic surgeon shows that our method could correctly segment the bones and fragments. (C) 2001 Elsevier Science B.V. All rights reserved.

This study was conducted to clarify the relationship between the input devices for the Internet and the factors such as age, health status, and future need of a mobile phone (MP) and the Internet through personal computer (IPC). The magnitude of the relation was evaluated by the odds ratio (OR). The survey was done in December 2000 and January 2001. The study sample was forty women, 78.5±4.5 years of age (70-89), who lived in their own houses in Y Town near Himeji City, Hyogo, Japan. As a result of analysis, the touch screen was the most preferred input device among the elderly with high age [OR=3.86, 95% confidence interval (CI) 0.83-18.98]. It was also preferred by subjects with more difficulties (OR=5.00, 95% CI 1.05-25.41) and less future need of a MP and the IPC (OR=7.22, 95% CI 1.34-43.88). These results suggest that the future need of a MP and the IPC was the most highly correlated with the input device for the Internet.

Hemispheric dominance for language has been assessed by the Wada test which is an invasive technique with considerable risk of complications. Recently, optical topography system has been used as a non-invasive method which is applicable in detecting the language dominance. Our method can be divided into three steps. Firstly, preprocessing is applied to the data. Secondly, we estimate the degree of the speech area. Finally, we calculate laterality index based on the degree of the speech area. Our results demonstrate that our method is effective for decision of language dominance.

This paper proposes an automated discriminating method of material using ultrasonic imaging that can conform change of tissue elasticity. The proposed method employs fuzzy infernce. Firstly, three attributes are acquired from ultrasonic wave of the known material. Secondly, membership functions are constructed using three attributes automatically. Finally, we presume the material using the function from unknown ultrasonic wave. We apply this method to two situations with bone phantom and without bone phantom. The experimental results on six materials showed that each presumed rates under the two situations were 86.6% and 63.3%.

This paper shows an automated method for segmenting the lateral ventricles from human brain MR images. It enables us to automatically volumetry and construct 3-D rendering images. First of all we segment the whole brain and the cerebrospinal fluid (CSF) by using a 3-D MR image processing software developed by Hata et at Our method for segmenting the lateral ventricles is based on fuzzy inference techniques, which is able to represent expert's knowledge and to introduce the knowledge to image processing The employed knowledge on the lateral ventricles is the location, the intensity, and the shape The proposed method was applied to MR volumes of 20 normal volunteers, 20 Alzheimer disease and 20 hydrocephalus patients The experimental results validated that this method was able to segment the lateral ventricles with high accuracy.

© Springer-Verlag Berlin Heidelberg 2001. This paper describes a new clustering method based on rough set theory. This method classifies objects according to the indiscernibility relations defined on the basis of relative similarity. First, an initial equivalence relation, which evaluates local similarity of objects, is assigned to every object. Then modification of the initial equivalence relations is performed by examining global relationships among them. An initial equivalence relation will be modified if it gives excessively fine classification to the objects. Consequently, generation of small category is suppressed and adequately coarse clusters are formed. Experimental results on the artificial data showed that this method produced good clustering results for both of nominal and numerical data.

Volume visualization of cerebral blood vessels is highly significant for diagnosis of the cerebral diseases. It is because the automated segmentation of the blood vessels from an MR angiography (MRA) image is a knotty problem that there are few works on it. This paper proposes an automated method to segment the blood vessels from 3D time of flight (TOF) MRA volume data. The method consists of: (1) removing the background, (2) volume quantization by watershed segmentation, and (3) classification of primitives by using an artificial neural network (NN). In the proposed method, the NN classifies each primitive, which is a clump of voxels, by evaluating the intensity and the 3D shape. The method was applied to seven MRA data sets. The evaluation was done by comparing with the manual classification results. The average classification accuracy was 80.8%. The method also showed the volume visualizations using target maximum intensity projection (target MIP) and surface shaded display (SSD). The evaluation by a physician showed that unclear regions on the conventional image were clearly depicted on applying the method, and that the produced images were quite interesting for diagnosis of cerebral diseases such as aneurysm and encephaloma. The quantitative and qualitative evaluations showed that the method was appropriate for blood vessel segmentation.

This paper proposes a clustering method for nominal and numerical data based on Rough Sets and its application to knowledge discovery in the medical database. Classification is performed according to the indiscernibility relations defined on the basis of relative similarity between objects. The similarity is defined as a combination of two types of similarity measures: the Hamming distance for nominal attributes and the Mahalanobis distance for numerical attributes. Excessive generation of small category is suppressed by modifying similar equivalence relations into the same equivalence relation. An analysis of the meningoencephalitis diagnosis database was performed to validate this method. The result showed that this method could deal well with both types of attributes and discover the primary factors for diagnosis.

© Springer-Verlag Berlin Heidelberg 2001. This paper presents a clustering method for nominal and numerical data based on rough set theory. We represent relative simi- larity between objects as a weighted sum of two types of distances: the Hamming distance for nominal data and the Mahalanobis distance for numerical data. On assigning initial equivalence relations to every ob- ject, modification of slightly different equivalence relations is performed to suppress excessive generation of categories. The optimal clustering re- sult can be obtained by evaluating the cluster validity over all clusters generated with various values of similarity thresholds. After classifica- tion has been performed, features of each class are extracted based on the concept of value reduct. Experimental results on artificial data and amino acid data show that this method can deal well with both types of attributes.

This paper proposes a clustering method for nominal and numerical data based on Rough Sets and its application to knowledge discovery in the medical database. Classification is performed according to the indiscernibility relations defined on the basis of relative similarity between objects. The similarity is defined as a combination of two types of similarity measures: the Hamming distance for nominal attributes and the Mahalanobis distance for numerical attributes. Excessive generation of small category is suppressed by modifying similar equivalence relations into the same equivalence relation. An analysis of the meningoencephalitis diagnosis database was performed to validate this method. The result showed that this method could deal well with both types of attributes and discover the primary factors for diagnosis.

The world currently faces the new wave of industrial civilization, that is regarded as the changes caused by the Revolution of Information Technology (IT). This study is conducted to clarify how a mobile phone (MP) and the Internet through personal computer (IPC) have been actually used. We also examine what kind of interface (input device) is preferred by the elderly, by using a questionnaire regarding the IT equipment. This study adopts not only the χ2 test but also the rough set analysis. As a result of the rough set analysis and χ2 test, those who were interested in a MP preferred, Japanese kana syllabary (JKS) the most. The rough set analysis also discovered the facts that those interested in the keyboard had a preference for the IPC and also that those interested in the JKS and the speech preferred a MP. We can conclude that the complementary use of the rough set analysis and χ2 test bring more powerful knowledge discovery. © Springer-Verlag 2001.

This paper presents a computer-aided diagnosis (CAD) system, which be able to segment the whole brain, the brain portions, the cerebrospinal fluid (CSF) and the lateral ventricles from human brain MR images, and then to give the volumes and the 3-D volume renderings. In the system, the whole brain and the brain portions are automatically segmented by using previously developed software (Hata et al., IEEE SMC-C, 2000). In this paper, we introduce a new concept of a representative line to segment the CSF and the lateral ventricles. The representative line is automatically detected by inferring its direction with evaluating the location in the head, the position in the CSF, and the shape of the line. The inference is preformed with fuzzy inference technique. The developed system was applied to MR volumes of 20 normal subjects, 20 Alzheimer disease and 20 hydrocephalus patients. The segmentation error ratio was 1.98% in comparison with the volumes of manually delineated region. This allows us to experimentally characterize modes of variation that are indicative of disease processes.

This paper proposes a computer aided diagnosis system of meniscal tears from 3D human knee MR image with T1-weighted and T2-weighted MR images. The first step of our method is a 3D image registration between both images on the computer display by manual. The second step determines the candidate region of the menisci from T2-weighted MR image aided by fuzzy if-then rules with respect to the location and the intensity. The final step determines the voxels in the menisci from the candidate region by using T1-weighted and T2-weighted MR images. We applied this method to several subjects. All voxels in the menisci of each subject were successfully identified and their 3D surfaces were displayed. Thus, our developed system would improve to diagnose the meniscal tears. © Springer-Verlag 2001.

Surface rendering of the endorrhachis is one of the effective techniques to diagnose diseases of the endorrhachis caused by spine malformations. It requires a procedure for segmenting the 3D endorrhachis from many MR images consisting the whole human body, however, few automated extraction methods have been developed before. In this paper, we propose a fuzzy rule-based approach for extracting the endorrhachis from MR lumbar images. Our target portion, the endorrhachis, consists of the spinal cord and the nerve roots. This method is based on knowledge of location and knowledge of intensity. They are expressed by fuzzy if-then rules and compiled to a total degree as the measure of extraction. Our extraction method can be divided into two steps. Firstly, the method extracts the spinal cord by using thresholding and labeling. Secondly, our method extracts the nerve roots with the knowledge derived from anatomical information. The experimental results showed that our method successfully extracted the target portion, and that the results were useful to diagnose the compression of the endorrhachis.

This paper proposes the new method for suppressing the inherent wave of the sensor system. The inherent wave often prevents from analyzing the signal of target. Previous method calculated the inherent wave to solve this problem using two assumptions. However, the calculated wave is differ from the original inherent wave in unsatisfying the assumptions. Therefore, the new method estimates the wave using fuzzy-logic without the assumptions. As a result, this method was able to detect with higher accuracy than previous method. © Springer-Verlag 2001.

Functional magnetic resonance imaging (fMRI) provides high-resolution datasets that allow neuro-scientists to sensitively detect activated region relating to a given task. Cerebral hemodynamic response (HR) of neural activity at the activated-region is delayed and dispersed in time. This study proposes a new data-driven analysis for detecting the activated region and measuring the temporal delay of HRs. Our method can be divided into four steps. Firstly, a standard preprocessing (realigning and high pass filtering) is applied to the fMRI data. Secondly, we reconstruct the time course of signal in order to improve the time resolution. Thirdly, we derive the form and the delays of HRs using continuous wavelet transform. Finally, we give wavelet coefficient as the degree of activation. The proposed method has been applied to 7 right-handed subjects on right or left hand-gripping task. Experimental results showed that primary motor cortex (M1) and the supplementary motor area (SMA) were detected as activated region, and means of temporal delays of HRs were right/left = 4.49/4.92 sec at SMA and 5.41/5.46 sec at M1. Consequently, our method could classify the difference of HR delay between M1 and SMA.

This paper presents a rough-sets-based method for clustering data composed of both nominal and numerical attributes. Similarity between objects is determined as a weighted sum of their Hamming / Mahalanobis distances. The method then forms equivalence relations among objects based on their relative similarity, and classifies them into some categories according to the relations. Similar equivalence relations will be merged into an equivalence relation so that it represents more simple knowledge which generates adequate number of categories. The optimal clustering result can be obtained by evaluating the cluster validity, defined using upper and lower approximations of a cluster, over all clusters generated with various values of similarity thresholds. Experimental results on the amino acid data show that our method can deal well with both attributes and produces good clustering results.

Near-infrared spectroscopy (NIRS) is a recently developed method, which can investigate the human brain function with noninvasive, high time resolution, and high portability. However, there are few discussions on post-processing of time series data taken by the NIRS because of the difficulty of understanding the obtained data and the complexity of the human higher-order brain function. This paper discusses on an analysis of such a time series. The analysis method is based on fuzzy c-means (FCM) clustering and wavelet transform, and it divides the time series of a measurement point into some clusters with respect to wavelet coefficients. To evaluate the performance of the method, it has been applied to four healthy volunteers, and three brain-dead patients. The results showed that the proposed method could segment the NIRS time series into some clusters that may represent brain states, and could estimate the number of clusters. © Springer-Verlag 2001.

Surface rendering of the endorrhachis is one of the effective techniques to diagnose diseases of the endorrhachis caused by spine malformations. It requires a procedure for segmenting the 3D endorrhachis from many MR images consisting the whole human body, however, few automated extraction methods have been developed before. In this paper, we propose a fuzzy rule-based approach for extracting the endorrhachis from MR lumbar images. Our target portion, the endorrhachis, consists of the spinal cord and the nerve roots. This method is based on knowledge of location and knowledge of intensity. They are expressed by fuzzy if-then rules and compiled to a total degree as the measure of extraction. Our extraction method can be divided into two steps. Firstly, the method extracts the spinal cord by using thresholding and labeling. Secondly, our method extracts the nerve roots with the knowledge derived from anatomical information. The experimental results showed that our method successfully extracted the target portion, and that the results were useful to diagnose the compression of the endorrhachis.

This paper proposes a computer aided diagnosis system of meniscal tears from 3D human knee MR image with T1-weighted and T2-weighted MR images. The first step of our method is a 3D image registration between both images on the computer display by manual. The second step determines the candidate region of the menisci from T2-weighted MR image aided by fuzzy if then rules with respect to the location and the intensity. The final step determines the voxels in the menisci from the candidate region by using T1-weighted and T2-weighted MR images. We applied this method to several subjects. All voxels in the menisci of each subject were successfully identified and their 3D surfaces were displayed. Thus, our developed system would improve to diagnose the meniscal tears.

To enhance fault tolerance ability of the feedforward neural networks (NNs for short) implemented in hardware, we discuss the learning algorithm that converges without adding extra neurons and a large amount of extra learning time and cycles. Our algorithm modified from the standard backpropagation algorithm (SBPA for short) limits synaptic weights of neurons in range during learning phase. The upper and lower bounds of the weights are calculated according to the average and standard deviation of them. Then our algorithm reupdates any weight beyond the calculated range to the upper or lower bound. Since the above enables us to decrease the standard deviation of the weights, it is useful in enhancing fault tolerance. We apply NNs trained with other algorithms and our one to a character recognition problem. It is shown that our one is superior to other ones in reliability, extra learning time and/or extra learning cycles. Besides we clarify that our algorithm never degrades the generalization ability of Ns although it coerces the weights within the calculated range.

The diameter of the yarn has been defined in the relationship between length and weight of the yarn. However, the diameter of the yarn is different by the convergence condition of the fiber that constitutes the yarn. Then, the new filter is made for the purpose of measuring the appearance diameter of the yarn. In this study, weighting filter to the differential image got by dividing the differential histogram values by the lightness histogram values was made. The measurement method of the appearance diameter of the yarn using this filter was proposed.