Semi-supervised learning (SSL) has shown strong potential in reducing the reliance on large-scale voxel-level annotations for 3D medical image segmentation. However, existing SSL methods often suffer from unstable training and limited generalization due to unreliable pseudo-labels and insufficient structural modeling in unlabeled data. These challenges are especially evident in volumetric contexts, where anatomical structures exhibit high inter-class imbalance and complex spatial dependencies. To address these issues, we propose a semi-supervised framework built upon a single-network architecture that integrates feature learning, consistency regularization, and pseudo-label reliability modeling in a unified manner. The framework comprises three key components: 1) a Confidence-aware Multi-level Fusion Network (CMFN) for capturing robust multi-scale semantic representations; 2) a Semantic-Enhanced Center Alignment (SECA) module to align feature distributions of group-level anatomical structures and mitigate semantic drift in pseudo-labels; and 3) a Group-Guided Reliability Assessment (GGRA) module that enhances pseudo-label reliability by modeling confidence errors in a group-aware structural context. Together, these modules enhance both feature discriminability and the reliability of pseudo-labels.We evaluate our framework on three public 3D medical image segmentation benchmarks: LA, BTCV, and BraTS19. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art approaches under limited annotation, achieving superior accuracy and generalization across diverse anatomical structures and segmentation tasks.

The exploitation of label hierarchy is crucial for effective brain tumor segmentation. Nevertheless, existing methods grapple with two key limitations. First, they lack the hierarchical dependency of predictions across different label levels, rendering the network outputs less interpretable. Second, they fail to exploit the hierarchical similarity among labels, thus hindering potential accuracy enhancement. To address these limitations, we present a novel framework termed deep hierarchy-aware segmentation (DHAS), which achieves both hierarchically interpretable and high-accuracy predictions. Specifically, to generate hierarchical predictions, the network is designed to output pixel-wise probability conditional upon the parent label and is hybrid-trained from conditional to unconditional probability. To utilize the label similarity, we propose a tree-triplet loss, which imposes the hierarchy-induced distance within the feature embedding space. Experimental results on three datasets, BraTS2018, BraTS2019 and BraTS2020, show that our proposed framework achieves significantly better performance than other hierarchy-exploiting methods, and it ranks fifth top among 383 participating methods in Brats2020 Challenge. The improved performance and interpretable predictions promise the potential of DHAS for clinical applications in brain tumor segmentation. Furthermore, its generalization is demonstrated for cardiac segmentation on ACDC dataset.

Hierarchical approaches have been tremendously successful at multi-label segmentation. However, it has been shown they may seriously suffer from the problem of only imposing constraints on shallow layers while ignoring deep relationships in the label space. In this paper we overcome this limitation through a hierarchical multi-class group correlation learning (HMGC). Thus, we first transform regional constraints into voxel vector correlations in a high-dimensional space. After performing transformation, we compute a voxel vector correlation matrix to group voxel vectors to reduce disparities between erroneous and valid vectors. We then introduce two loss functions: intra-class group loss, which minimizes differences within the same class, and inter-class group loss, which adjusts distances between class group centers and voxel vectors. This, in turn, can be used to mitigate bias propagation and improve segmentation accuracy. The effectiveness of our method is demonstrated on three Brain Tumor Segmentation Challenge datasets: BraTS2018, BraTS2019, and BraTS2020. Moreover, generalization of our method is evaluated on the ACDC MICCAI'17 Challenge Dataset. Our HMGC model ranks first in overall score on Brats2020 and achieves one of the most competitive results in cardiac segmentation.

Deep neural network (DNN) approaches have shown remarkable progress in automatic Chest X-rays classification. However, existing methods use a training scheme that simultaneously trains all abnormalities without considering their learning priority. Inspired by the clinical practice of radiologists progressively recognizing more abnormalities and the observation that existing curriculum learning (CL) methods based on image difficulty may not be suitable for disease diagnosis, we propose a novel CL paradigm, named multi-label local to global (ML-LGL). This approach iteratively trains DNN models on gradually increasing abnormalities within the dataset, i,e, from fewer abnormalities (local) to more ones (global). At each iteration, we first build the local category by adding high-priority abnormalities for training, and the abnormality's priority is determined by our three proposed clinical knowledge-leveraged selection functions. Then, images containing abnormalities in the local category are gathered to form a new training set. The model is lastly trained on this set using a dynamic loss. Additionally, we demonstrate the superiority of ML-LGL from the perspective of the model's initial stability during training. Experimental results on three open-source datasets, PLCO, ChestX-ray14 and CheXpert show that our proposed learning paradigm outperforms baselines and achieves comparable results to state-of-the-art methods. The improved performance promises potential applications in multi-label Chest X-ray classification.

Due to bone deformation and joint space narrowing in diseased hips, accurate segmentation for pelvis, and femur from hip computed tomography (CT) images remains a challenging task. Therefore, the paper presents a fully automatic segmentation framework for the pelvis and femur in both of healthy and diseased hips. The framework involves three steps: preprocessing, coarse segmentation, and refinement. It starts with a preprocessing procedure to extract the volume of interest (VOI) from original CT images. Then, a coarse segmentation of bone has been obtained by classifying the VOI as bone and nonbone parts based on conditional random field (CRF) model. Finally, the bone is further divided into the pelvis and femur using a patch-based refinement method. The innovation of this study is the novel patch-based refinement method that is particularly suitable for diseased hips. The refinement method starts from the boundary of coarse segmentation, and propagates to the neighbors only when the label is not consistent with the label of CRF-based classification, it increases the reliability of segmentation for diseased hips with bone deformation. We incorporate neighborhood information to label fusion so that final label estimation is more accurate and robust for diseased hips with joint space narrowing. In total, 60 CT data sets, which included 78 healthy hemi-hips and 42 diseased hemi-hips, were used, and three-fold cross validations were carried out. Compared to two state-of-the-art methods, our method achieved significantly increased segmentation accuracy for the diseased hemi-hips, and is, therefore, more suited for automatic segmentation of diseased hips.

Neurons behave like transistors, but have fluctuating characteristics. In this paper, we show that several asynchronous multiplex communication channels can be established in a 2-D mesh neural network with randomly generated weights between eight neighbors. Neurons were simulated by integrate-and-fire neuron models without leakage and with fluctuating refractory period and output delay. If one of the transmitting neuron groups is stimulated, the signal is propagated in the form of spike waves. The corresponding receiving neuron group is able to identify the signal after having learned to form an asynchronous multiplex communication channel. The channel is composed of many intermediate/interstitial neurons working as relays. Each neuron can work as an I/O and as a relay element, i.e., as a multiuse unit. Grouping and synchronic firing is often seen in natural neuronal networks and seems to be effective for stable/robust communication in conjunction with spatial multiplex communication. This communication pattern corresponds to our wet lab experiments on cultured neuronal networks and is similar to sound identification by the ear and mobile adaptive communication systems.

We describe an active contour framework with accurate shape and size constraints on the vessel cross-sectional planes to produce the vessel segmentation. It starts with a multiscale vessel axis tracing in a 3D computed tomography (CT) data, followed by vessel boundary delineation on the cross-sectional planes derived from the extracted axis. The vessel boundary surface is deformed under constrained movements on the cross sections and is voxelized to produce the final vascular segmentation. The novelty of this paper lies in the accurate contour point detection of thin vessels based on the CT scanning model, in the efficient implementation of missing contour points in the problematic regions and in the active contour model with accurate shape and size constraints. The main advantage of our framework is that it avoids disconnected and incomplete segmentation of the vessels in the problematic regions that contain touching vessels (vessels in close proximity to each other), diseased portions (pathologic structure attached to a vessel), and thin vessels. It is particularly suitable for accurate segmentation of thin and low contrast vessels. Our method is evaluated and demonstrated on CT data sets from our partner site, and its results are compared with three related methods. Our method is also tested on two publicly available databases and its results are compared with the recently published method. The applicability of the proposed method to some challenging clinical problems, the segmentation of the vessels in the problematic regions, is demonstrated with good results on both quantitative and qualitative experimentations; our segmentation algorithm can delineate vessel boundaries that have level of variability similar to those obtained manually.

This paper describes a ultrasound (3-D US) system that aims to achieve augmented reality (AR) visualization during laparoscopic surgery, especially for the liver. To acquire 3-D US data of the liver, the tip of a laparoscopic ultrasound probe is tracked inside the abdominal cavity using a magnetic tracker. The accuracy of magnetic trackers, however, is greatly affected by magnetic field distortion that results from the close proximity of metal objects and electronic equipment, which is usually unavoidable in the operating room. In this paper, we describe a calibration method for intraoperative magnetic distortion that can be applied to laparoscopic 3-D US data acquisition; we evaluate the accuracy and feasibility of the method by in vitro and in vivo experiments. Although calibration data can be acquired freehand using a magneto-optic hybrid tracker, there are two problems associated with this method--error caused by the time delay between measurements of the optical and magnetic trackers, and instability of the calibration accuracy that results from the uniformity and density of calibration data. A temporal calibration procedure is developed to estimate the time delay, which is then integrated into the calibration, and a distortion model is formulated by zeroth-degree to fourth-degree polynomial fitting to the calibration data. In the in vivo experiment using a pig, the positional error caused by magnetic distortion was reduced from 44.1 to 2.9 mm. The standard deviation of corrected target positions was less than 1.0 mm. Freehand acquisition of calibration data was performed smoothly using a magneto-optic hybrid sampling tool through a trocar under guidance by realtime 3-D monitoring of the tool trajectory; data acquisition time was less than 2 min. The present study suggests that our proposed method could correct for magnetic field distortion inside the patient's abdomen during a laparoscopic procedure within a clinically permissible period of time, as well as enabling an accurate 3-D US reconstruction to be obtained that can be superimposed onto live endoscopic images.

This paper presents a framework for a cluster system that is suited for high-resolution image processing over the Internet during surgery. The system realizes high-performance computing (HPC) assisted surgery, which allows surgeons to utilize HPC resources remote from the operating room. One application available in the system is an intraoperative estimator for the range of motion (ROM) adjustment in total hip replacement (THR) surgery. In order to perform this computation-intensive estimation during surgery, we parallelize the ROM estimator on a cluster of 64 PCs, each with two CPUs. Acceleration techniques such as dynamic load balancing and data compression methods are incorporated into the system. The system also provides a remote-access service over the Internet with a secure execution environment. We applied the system to an actual THR surgery performed at Osaka University Hospital and confirmed that it realizes intraoperative ROM estimation without degrading the resolution of images and limiting the area for estimations.

Two-dimensional (2-D)/three-dimensional (3-D) registration techniques using single-plane fluoroscopy are highly important for analyzing 3-D kinematics in applications such as total knee arthroplasty (TKA) implants. The accuracy of single-plane fluoroscopy-based techniques in the determination of translation perpendicular to the image plane (depth position), however, is relatively poor because a change in the depth position causes only small changes in the 2-D silhouette. Accuracies achieved in depth position using conventional 2-D/3-D registration techniques are insufficient for clinical applications. Therefore, we propose a technique for improving the accuracy of depth position determination in order to develop a system for analyzing knee kinematics over the full six degrees of freedom (6 DOF) using single-plane fluoroscopy. In preliminary experiments, the behaviors of errors for each free variable were quantified as evaluation curves by examining changes in cost function with variations in the free variable. The evaluation curve for depth position was more jagged, and the curve peak less pointy, compared to the evaluation curves of the other five variables, and the curve was found to behave differently. Depth position is therefore optimized independently of the other variables, using an approximate evaluation curve of depth position prepared after initial registration. Accuracy of the proposed technique was evaluated by computer simulation and in vitro tests, with validation of absolute position and orientation performed for each knee component. In computer simulation tests, root-mean-square error (RMSE) in depth position was improved from 2.6 mm (conventional) to 0.9 mm (proposed), whereas for in vitro tests, RMSE improved from 3.2 mm to 1.4 mm. Accuracy of the estimation of the remaining two translational and three rotational variables was found to be almost the same as that obtained by conventional techniques. Results of in vivo tests are also described in which the possibility of full 6 DOF kinematic analysis of TKA implants is shown.

This paper describes several new methods and software for automatic segmentation of the pelvis and the femur, based on clinically obtained multislice computed tomography (CT) data. The hip joint is composed of the acetabulum, cavity of the pelvic bone, and the femoral head. In vivo CT data sets of 60 actual patients were used in the study. The 120 (60 x 2) hip joints in the data sets were divided into four groups according to several key features for segmentation. Conventional techniques for classification of bony tissues were first employed to distinguish the pelvis and the femur from other CT tissue images in the hip joint. Automatic techniques were developed to extract the boundary between the acetabulum and the femoral head. An automatic method was built up to manage the segmentation task according to image intensity of bone tissues, size, center, shape of the femoral heads, and other characters. The processing scheme consisted of the following five steps: 1) preprocessing, including resampling 3-D CT data by a modified Sinc interpolation to create isotropic volume and to avoid Gibbs ringing, and smoothing the resulting images by a 3-D Gaussian filter; 2) detecting bone tissues from CT images by conventional techniques including histogram-based thresholding and binary morphological operations; 3) estimating initial boundary of the femoral head and the joint space between the acetabulum and the femoral head by a new approach utilizing the constraints of the greater trochanter and the shapes of the femoral head; 4) enhancing the joint space by a Hessian filter; and 5) refining the rough boundary obtained in step 3) by a moving disk technique and the filtered images obtained in step 4). The above method was implemented in a Microsoft Windows software package and the resulting software is freely available on the Internet. The feasibility of this method was tested on the data sets of 60 clinical cases (5000 CT images).

Measuring the thickness of sheet-like thin anatomical structures, such as articular cartilage and brain cortex, in three-dimensional (3-D) magnetic resonance (MR) images is an important diagnostic procedure. This paper investigates the fundamental limits on the accuracy of thickness determination in MR images. We defined thickness here as the distance between the two sides of boundaries measured at the subvoxel resolution, which are the zero-crossings of the second directional derivatives combined with Gaussian blurring along the normal directions of the sheet surface. Based on MR imaging and computer postprocessing parameters, characteristics for the accuracy of thickness determination were derived by a theoretical simulation. We especially focused on the effects of voxel anisotropy in MR imaging with variable orientation of sheet-like structure. Improved and stable accuracy features were observed when the standard deviation of Gaussian blurring combined with thickness determination processes was around square root of 2/2 times as large as the pixel size. The relation between voxel anisotropy in MR imaging and the range of sheet normal orientation within which acceptable accuracy is attainable was also clarified, based on the dependences of voxel anisotropy and the sheet normal orientation obtained by numerical simulations. Finally, in vitro experiments were conducted using an acrylic plate phantom and a resected femoral head to validate the results of theoretical simulation. The simulated thickness was demonstrated to be well-correlated with the actual in vitro thickness.

This paper reports on the clinical application of a system for recovering the time-varying three-dimensional (3-D) left-ventricular (LV) shape from multiview X-ray cineangiocardiograms. Considering that X-ray cineangiocardiography is still commonly employed in clinical cardiology and computational costs for 3-D recovery and visualization are rapidly decreasing, it is meaningful to develop a clinically applicable system for 3-D LV shape recovery from X-ray cineangiocardiograms. The system is based on a previously reported closed-surface method of shape recovery from two-dimensional occluding contours with multiple views. To apply the method to "real" LV cineangiocardiograms, user-interactive systems were implemented for preprocessing, including detection of LV contours, calibration of the imaging geometry, and setting of the LV model coordinate system. The results for three real LV angiographic image sequences are presented, two with fixed multiple views (using supplementary angiography) and one with rotating views. 3-D reconstructions utilizing different numbers of views were compared and evaluated in terms of contours manually traced by an experienced radiologist. The performance of the preprocesses was also evaluated, and the effects of variations in user-specified parameters on the final 3-D reconstruction results were shown to be sufficiently small. These experimental results demonstrate the potential usefulness of combining multiple views for 3-D recovery from "real" LV cineangiocardiograms.

We propose a model to generate a group of artificial lives capable of coping with various environment which is a set of tasks, and show play or hobbies are necessary for the group of individuals to maintain the ability to cope with various changes of the environment as a whole. This may be an another aspect of the wide variety of essential group capabilities. If the variety in a species decreases it will become extinct. We show some simulation results: in a world where a greater variety of abilities are demanded in play the performance of the whole world calculated only from job tasks becomes steady and avoids the risk of extinction of the species. This is the effect of play.

The goal of this paper is to present appropriate line segmentation for intersections (X-junctions) and branches (T-junctions). In the local regions of intersections and branches, multiple orientations occur. A novel representation called "orientation space" is proposed, which is derived by adding the orientation axis to the abscissa and the ordinate of the image. The orientation space representation is constructed by treating the orientation parameter, to which Gabor filters can be tuned, as a continuous variable. The problem of multiple orientation line segmentation is dealt with by thresholding 3D images of the orientation space and then detecting the connected components therein. In this way, X-junctions and T-junctions are able to be separated effectively. Experimental results are presented using synthesized and real biomedical images. In particular, overlapping vessels in an x-ray coronary angiogram were well segmented by orientation space filtering.

A method of evaluating brain function using the metacomputer concept of the Globus system combined with a message-passing interface is described. The proposed method has the ability to exploit various geographically distributed resources and parallel computing linked to a high-technology medical instrumentation system, magnetoencephalography, to analyze the functional state of the brain. It is envisaged that the method will lead to the realization of an efficient telemedicine system for health care.

This paper describes a novel approach to tissue classification using three-dimensional (3D) derivative features in the volume rendering pipeline. In conventional tissue classification for a scalar volume, tissues of interest are characterized by an opacity transfer function defined as a one-dimensional (1D) function of the original volume intensity. To overcome the limitations inherent in conventional 1D opacity functions, we propose a tissue classification method that employs a multidimensional opacity function, which is a function of the 3D derivative features calculated from a scalar volume as well as the volume intensity. Tissues of interest are characterized by explicitly defined classification rules based on 3D filter responses highlighting local structures, such as edge, sheet, line, and blob, which typically correspond to tissue boundaries, cortices, vessels, and nodules, respectively, in medical volume data. The 3D local structure filters are formulated using the gradient vector and Hessian matrix of the volume intensity function combined with isotropic Gaussian blurring. These filter responses and the original intensity define a multidimensional feature space in which multichannel tissue classification strategies are designed. The usefulness of the proposed method is demonstrated by comparisons with conventional single-channel classification using both synthesized data and clinical data acquired with CT (computed tomography) and MRI (magnetic resonance imaging) scanners. The improvement in image quality obtained using multichannel classification is confirmed by evaluating the contrast and contrast-to-noise ratio in the resultant volume-rendered images with variable opacity values.

This paper describes the usefulness of computer assistance in the acquisition of "good" images for stenosis diagnosis and quantification in coronary angiography. The system recommends the optimal viewpoints from which stenotic lesions can be observed clearly based on images obtained from initial viewpoints. First, the viewpoint dependency of the apparent severity of a stenotic lesion is experimentally analyzed using software phantoms in order to show the seriousness of the problem. The implementation of the viewpoint determination system is then described. The system provides good user-interactive tools for the semiautomated estimation of the orientation and diameter of stenotic segments and the three-dimensional (3-D) reconstruction of vessel structures. Using these tools, viewpoints that will not give rise to foreshortening and vessel overlap can be efficiently determined. Experiments using real coronary angiograms show the system to be capable of the reliable diagnosis and quantification of stenosis.

This paper describes augmented reality visualization for the guidance of breast-conservative cancer surgery using ultrasonic images acquired in the operating room just before surgical resection. By combining an optical three-dimensional (3-D) position sensor, the position and orientation of each ultrasonic cross section are precisely measured to reconstruct geometrically accurate 3-D tumor models from the acquired ultrasonic images. Similarly, the 3-D position and orientation of a video camera are obtained to integrate video and ultrasonic images in a geometrically accurate manner. Superimposing the 3-D tumor models onto live video images of the patient's breast enables the surgeon to perceive the exact 3-D position of the tumor, including irregular cancer invasions which cannot be perceived by touch, as if it were visible through the breast skin. Using the resultant visualization, the surgeon can determine the region for surgical resection in a more objective and accurate manner, thereby minimizing the risk of a relapse and maximizing breast conservation. The system was shown to be effective in experiments using phantom and clinical data.

We propose a new model and new algorithm for acquiring symbolic description from a flow field. Our model is derived from physical data, and our algorithm is based on holomorphic complex function theory. Application of this method enables both qualitative and quantitative information to be obtained from a flow field. We also demonstrate the robustness of our method using a simulated flow field and a real image sequence.

This paper describes a method for the accurate recovery of time-varying 3D shapes with known cycle from images with different viewpoints as well as times; aiming at the recovery of the left ventricular shapes. Our recovery method is based on the integration of apparent contours from different viewpoints. We perform direct fitting to a 4D closed surface model based on B-splines so as to deal with fragmented contours such as extracted from X-ray cineangiocardiograms. The method is quantitatively evaluated using synthesized and real image sequences.

Neural-network theorems state that only when there are infinitely many hidden units is a four-layered feedforward neural network equivalent to a three-layered feedforward neural network. In actual applications, however, the use of infinitely many hidden units is impractical. Therefore, studies should focus on the capabilities of a neural network with a finite number of hidden units, In this paper, a proof is given showing that a three-layered feedforward network with N-1 hidden units can give any N input-target relations exactly. Based on results of the proof, a four-layered network is constructed and is found to give any N input-target relations with a negligibly small error using only (N/2)+3 hidden units. This shows that a four-layered feedforward network is superior to a three-layered feedforward network in terms of the number of parameters needed for the training data.

Penumbral imaging is a technique for imaging of neutrons or other penetrating radiations. The technique uses the facts that spatial information can be recovered from the shadow or penumbra that an unknown source casts through a simple large circular aperture. The limitation is that the straightforward image reconstruction will introduce some significant distortion for a large field of view because of nonisoplanaticity of the aperture point spread function. A genetic algorithm (GA) is proposed for reconstruction of penumbral images, and the technique allows distortion-free reconstruction over a large field of view. Furthermore, because in GA the complicated a priori constraints can be easily incorporated by the appropriate modification of the cost function, the algorithm is also tolerant of the noise.

A parallel genetic algorithm based on the island model for image restoration is presented. The algorithm divides a large population into smaller subpopulations and executes the main loop of the traditional genetic algorithm on each processor with its own subpopulation in parallel. Its performance is evaluated in a multi-workstation environment. The simulation results show that the algorithm achieves a linear speed-up with the number of processors. The parallel algorithm is also shown to have better performance on image restoration than the traditional genetic algorithm.

A post-processing technique has been developed to suppress the magnetic resonance imaging (MRI) artifact arising from object planar rigid motion. In two-dimensional Fourier transform (2-DFT) MRI, rotational and translational motions of the target during magnetic resonance magnetic resonance (MR) scan respectively impose nonuniform sampling and a phase error an the collected MRI signal. The artifact correction method introduced considers the following three conditions: (1) for planar rigid motion with known parameters, a reconstruction algorithm based on bilinear interpolation and the super-position method is employed to remove the MRI artifact, (2) for planar rigid motion with known rotation angle and unknown translational motion (including an unknown rotation center), first, a super-position bilinear interpolation algorithm is used to eliminate artifact due to rotation about the center of the imaging plane, following which a phase correction algorithm is applied to reduce the remaining phase error of the MRI signal, and (3) to estimate unknown parameters of a rigid motion, a minimum energy method is proposed which utilizes the fact that planar rigid motion increases the measured energy of an ideal MR image outside the boundary of the imaging object; by using this property all unknown parameters of a typical rigid motion are accurately estimated in the presence of noise. To confirm the feasibility of employing the proposed method in a clinical setting, the technique was used to reduce unknown rigid motion artifact arising from the head movements of two volunteers.

A computer postprocessing technique is developed to remove MRI artifact arising from unknown translational motion in the imaging plane. Based on previous artifact correction methods, the improved technique uses two successive steps to reduce read out and phase-encoding direction artifacts: First, the spectrum shift method is applied to remove read-out axis translational motion. Then, the phase retrieval method is employed to eliminate the remaining subpixel motion of the read-out axis and the entire motion of the phase-encoding axis. In the presence of noise, to protect edge detection (in the spectrum shift method), two high-density gray-level markers are added, one to each side of the imaging object. Experimental results with an actual MR scan confirmed the ability of the method to correct the artifact of an MR image caused by unknown translational motion in the imaging plane.

We make a comparision of classification ability between BPN (BackPropagation Neural Network) and k-NN (k-Nearest Neighbor) classification methods. Voice data and patellar subluxation images are used. The result was that the average recognition rate of BPN was 9.2 percent higher than that of the k-NN classification method. Although k-NN classification is simple in theory, classification time was fairly long. Therefore, it seems that real time recognition is difficult. On the other hand, the BPN method has a long learning time but a very short recognition time. Especially if the number of dimensions of the samples is large, it can be said that BPN is better than k-NN in classification ability.

Three-dimensional reconstruction from a perspective 2D image using mirrors is addressed. The mirrors are used to form symmetrical relations between the direct image and mirror images. By finding correspondences between them, the 3D shape can be reconstructed by means of plane symmetry recovering method using the vanishing point. Two constraints are used in determining the correspondence. In the case where only one mirror is used, invisible parts both in the direct image and in the mirror image may still remain. Using multiple mirrors, however, occluded parts will decrease or disappear, and occlusion-free object reconstruction becomes possible.>

A phoneme filter neural network (PFN) approach to vowel recognition is described. The PFN is a multilayer neural network with fewer hidden units than input units prepared for each of the phoneme categories. Each network is trained as identity mapping by speech data belonging to one phoneme category. In the recognition process, the similarity between the input data and output data is computed for each network. The results of an experiment involving the Japanese vowel recognition task showed that the PFN recognition rates for the top two or more choices are higher than those of a conventional three-layer neural network and the PFN outputs represented candidate likelihoods. It was also confirmed that the PFN has a mapping ability and recognition performance superior to those of the linear K-L transformation method because of the nonlinearity of the PFN.>

Two experiments aimed at improving the performance of a noise reduction neural network are described. One approach is to split the last single affine transformation of the noise reduction neural network into two affine transformations and to adaptively control these transformations for noise components and speech components. The other is to further split the last single affine transformation to yield 22 affine transformations tuned for abstract concepts, or phonemes, and to adaptively control these transformations according to certain selection criteria. The latter is a refinement of the former. Both approaches are based on the fact that the speech component is more easily separated form the noise component in the second hidden layer output than in the physical input space of the network.>

An analysis of the four-layer, feedforward, noise-reduction neural network proposed by S. Tamura author and A. Waibel (Int. Conf. Acoust., Speech and Signal Proc., p.553-6, 1988) is described. Each layer has 60 units and is fully interconnected with the next higher layer. The input of the network is given by a 60-point-long (at 12-kHz sampling rate) noisy waveform, and the output is a 60-point-long noise-free waveform. The network was trained using the back-propagation learning algorithm. The network is divided into three subelements for the analysis. Each element stands for a transformation from a layer output to the next higher layer output. First, the transformation from the input layer to the first hidden layer is analyzed, showing that the linear part of the transformation performs linear noise reduction as well as linear speech/noise-characteristic extraction. It is also shown that the transformation from the first hidden layer to the second hidden layer greatly compresses the noise region of the first hidden layer by sigmoid nonlinearities, while preserving its speech region, and the transformation from the second hidden layer to the output layer linearly suppresses the noise components. The spectra of basis vectors spanning an output waveform space show poor higher formant structures.>

If an object to be observed is much larger than the field of vision, we should observe it part by part as in the case of eye fundus photographs or gastrocamera photographs. Subpictures obtained through partial observation should be assembled into a mosaic picture which shows the whole object. An arrangement graph which shows adjacency and spatial differences among subpictures is defined, and the properties of the arrangement graph are discussed. Furthermore, the condition of the arrangement graph for a mosaic picture to be defect-free, that for a subpicture to be redundant in a mosaic picture, and the way to find subpictures surrounding defect areas of a mosaic picture are shown. The discussion is based on a graph theory and a notion of space vectors.

An on-off sequence of an unknown signal pattern that is time-varying or fixed is dealt with. The optimum analogue feedback signal is obtained, which minimizes the average energy transmitted in the forward direction. The nonsupervised situation is also discussed. The effectiveness of the feedback method over the nonfeedback method is twice for the typical case.

A learning model of fuzzy formal language is proposed and discussed. We continue training the learning machine by giving sets of sentences sequentially. As a result of parsing of the given teaching sentences, the learning machine reinforces fuzzy grades of membership of productions in an inherent fuzzy grammar of the machine. The convergence of the proposed model is considered, and it is shown that the grades of membership of desired productions are intensified by choosing an adequate teaching sequence of the sentence set. Furthermore, a concept of ``strongly equivalent,'' in which two grammars are not distinguished by any teaching sequence, is introduced.

Methods of synchronization for an unknown signal sequence by learning procedure, which computes a posteriori probabilities for the signal location, are presented. First, to avoid the difficulties of the interference between successive intervals, a signal sequence with ample guard spaces for getting rid of the interference is treated. Next, the case without guard space is treated. To the nonsupervised case, some methods are applied. The methods proposed in this paper have a form of double learning with respect to the unknown signal form and the signal location. Also some results of computer simulation are presented.

Some recognizers for stochastic time-varying patterns with additive noise are studied. As in binary communication channels with fading, it is supposed that the fluctuation of a pattern (or signal) may be approximated by a stationary Gaussian autoregressive process with known parameters. Each measurement belongs to either of two classes: the pattern plus noise or noise alone. Under these assumptions, optimum dichotomizers with supervized learning are discussed. To the nonsupervised problems, the decision-directed approach and the modified-decision-directed approach are applied. Also some experimental results are presented.

A method of classifying patterns using fuzzy relations is described. To start with, we give a suitable value of the measure of subjective similarity to each pair of patterns that is taken from the population of patterns to be classified. Then a similitude between any two patterns is calculated by using the composition of a fuzzy relation. The similitude induces an equivalence relation. Consequently, we can classify the present population of the patterns into some classes by the equivalence relation. An experiment of the classification of portraits has been performed to test the method proposed here.