Catálogo de publicaciones - libros

Topological structure of Probability Neural Network (PNN) is usually complex when it is trained with large-scale and high-redundant training samples. Aiming this problem, PNN is analyzed and simplified by using probability calculation and multiplication formula. At first, input data of training samples was statistical analyzed by using Principal Component Analysis (PCA). PNN topological structure was optimized based on the statistical results. Subsequently, a complete learning algorithm was provided to avoid the artificial set of smoothing parameters. And Simulated Annealing (SA) coefficient was introduced to increase learning speed and stability. Eventually, the optimized PNN was applied to real problem. The test result validated that the optimized PNN had simpler structure and higher efficiency than typical PNN in the application with large-scale and high-redundant training samples.

With Kohonen’s self-organizing map based brain MRI image segmentation, there are still some regions which are not partitioned accurately, particularly in the transitional regions of gray matter and white matter, or cerebrospinal fluid and gray matter. In this paper, we propose a dynamically growing hierarchical self-organizing map integrated with a multi-scale feature vector to overcome the problem mentioned above, which uses the spatial relationships between image pixels and using multi-scale processing method to reduce the noise effect and the classification ambiguity. The efficacy of our approach is validated by extensive experiments using both simulated and real MRI images.

Detection of biomarkers capable of predicting a patient’s risk of major adverse cardiac events (MACE) is of clinical significance. Due to the high dynamic range of the protein concentration in human blood, applying proteomics techniques for protein profiling can generate large arrays of data for development of optimized clinical biomarker panels. The objective of this study is to discover an optimized subset of biomarkers for predicting risk of MACE containing less than ten biomarkers. In this paper, we connect linear SVM with PLS feature selection and extraction. A simplified PLS algorithm selects a subset of biomarkers and extracts latent variables and prediction performance of linear SVM is dramatically improved. The proposed method is compared with a widely used PLS-Logistic Discriminant solution and several other reported methods based on the MACE prediction experiments.

In this paper we propose a comparative study of Artificial Neural Networks (ANN) and Artificial Immune Systems. Artificial Immune Systems (AIS) represent a novel paradigm in the field of computational intelligence based on the mechanisms that allow vertebrate immune systems to face attacks from foreign agents (called antigens). Several similarities as well as differences have been shown by Dasgupta in [1]. Here we present a comparative study of these two approaches considering evolutions of the concepts of ANN and AIS, respectively hybrid neural systems, Artificial Immune Recognition Systems (AIRS) and aiNet. We tried to establish a comparison among these three methods using a well known dataset, namely the Wisconsin Breast Cancer Database. We observed interesting trends in systems’ performances and capabilities. Peculiarities of these systems have been analyzed, possible strength points and ideal contexts of application suggested. These and other considerations will be addressed in the rest of this manuscript.

Many different approaches for the modeling of neural networks were presented in the literature (e.g. [4]). Generally the object-oriented approach proved itself as most appropriate. It provides a concise but comprehensive framework for the design of neural networks in terms of its static and dynamic components, i.e. the information structure and its methods in the object-oriented notion.

This paper presents a framework for the conceptual and physical integration of neural networks into object-relational database systems. The static components comprise the structural parts of a neural network, as the neurons and connections, higher topological structures as layers, blocks and network systems. The dynamic components are the behavioral characteristics, as the creation, training and evaluation of the network. Finally the implementation of the new system based on the proposed framework is presented.

There are fewer effective methods to accurately discriminate the coronary microcirculatory dysfunction from the normal coronary microcirculation. Rather than traditional approaches only considering a single hemodynamic parameter, a novel scheme is proposed based on the generalized relevance learning vector quantization (GRLVQ) using multiple parameters (features). Naturally integrating the tasks of feature selection and classification, this scheme circularly adopts GRLVQ to gradually prune the unimportant features according to their weighting factors. In each circulation, the prototypes are generated for classification and the classification accuracy is obtained. Finally, the feature subset with the highest classification accuracy is selected and the corresponding classifier is also achieved. This approach not only simplifies the classifier but also enhances the classification performance. The method is verified on the physiological data collected from animals, and proved to be superior to the traditional single-parameter method.

How to locate the neural activation sources effectively and precisely from the magnetoencephalographic (MEG) recording is a critical issue for the clinical neurology and brain functions research. Multiple signal classification (MUSIC) algorithm and recursive MUSIC algorithm are widely used to locate multiple dipolar sources from the MEG data. The drawback of these algorithms is that they run very slowly when scanning a three-dimensional head volume globally. In order to solve this problem, a novel MEG sources localization scheme based on genetic algorithm (GA) is proposed. First, this scheme uses the property of global optimum of GA to estimate the rough source location. Then, combined with grids in small area, the accurate dipolar source localization is performed. Furthermore, we introduce the adaptive crossover and mutation probability, two-point crossover operator, periodical substitution and niche strategies to overcome the disadvantage of GA which falls into local optimum occasionally. Experimental results show that the proposed scheme can improve the speed of source localization greatly and its accuracy is satisfactory.

It is difficult to directly co-register the 3D FMT (Fluorescence Molecular Tomography) image of a small tumor in a mouse whose maximal diameter is only a few mm with a larger CT image of the entire animal that spans about ten cm. This paper proposes a new method to register 2D flat and projected CT image first to facilitate the registration between small 3D FMT images and large CT images. And a novel algorithm Bayesian Iterative Closest Point (BICP) is introduced and validated in 2D affine registration. The visualization of the alignment of the 3D FMT and CT image through 2D registration shows promising results that would lead to automated 3D registration.

Some foot plant pathologies, like cave and flat foot, are normally detected by a human expert by means of footprint images. Nevertheless, the lack of trained personal to accomplish such massive first screening detection efforts precludes the routinely diagnostic of the above mentioned pathologies. In this work an innovative automatic system for foot plant pathologies based on neural networks (NN) is presented. We propose the use of principal components analysis to reduce the number of inputs to the NN and therefore increasing the efficiency of the training algorithm. The results achieved with this system evidence the feasibility of establishing automatic diagnosis systems based on the footprint image. These systems are of a great value specially in apart areas and are also suited to carry on massive first screening health campaigns.

In this paper, we detect threshold-driven phase transitions in the homogeneous random neural network. When the neurons are arranged as one dimension, the critical threshold is two, while in two dimensions counterpart, the critical threshold is four. We declare that random neural network is a specific case of Abstract neural automata. So we conclude that phase transitions in the random neural network can produce thought in human brain. We successfully apply the network to interpret the relation between diseases and syndrome in Traditional Chinese Medicine.