ARTIFICIAL NEURAL NETWORKS IN MEDICINE AND BIOLOGY Center for Biomedical Informatics State University of Campinas, Campinas - Brazil Abstracts of published work by the Center Status of Aug 15 1992 ----------------------------------------------------------- A HIGH-LEVEL LANGUAGE AND MICROCOMPUTER PROGRAM FOR THE DESCRIPTION AND SIMULATION OF NEURAL ARCHITECTURES Sabbatini, RME and Arruda-Botelho, AG Center of Biomedical Informatics, Neurosciences Applications Group, State University of Campinas, Campinas, SP, BRAZIL) The description, representation and simulation of complex neural network structures by means of computers is an essential step in the investigation of model systems and inventions in the growing field of biological information processing and neurocomputing. The handcrafting of neural net architectures, however, is a long, tedious, difficult and error-prone process, which can be substituted satisfactorily by the neural network analogue of a computer program or formal symbolic language. Several attempts have been made to develop and apply such languages: P3, Hecht- Nielsen's AXON, and Rochester's ISCON are some recent examples. We present here a new tool for the formal description and simulation of artificial neural tissues in microcomputers. It is a network editor and simulator, called NEUROED, as well as a compiler for NEUROL, a high-level symbolic, structured language which allows the definition of the following elements of a neural tissue: a) elementary neural architectonic units: each unit has the same number of cells and the same internal interconnecting pattern and cell functional parameters; b) elementary cell types: each cell can be defined in terms of its basic functional parameters; synoptic interconnections inside an architectonic unit (axonic delay, weights and signal can be defined for each); a cell can fan out to several others, with the same synoptic properties; c) synaptic interconnections among units; d) cell types and architectonic units can be replicated automatically across neural tissue and interconnected; e) cell types and architectonic units can be named and arranged in hierarchical frames (parameter inheritance). NEUROED's underlying model processing element (PE) is a simplified Hodgkin-Huxley neuron, with RC-model, temporal- summation, passive electrotonic potentials at dendritic level, and a step transfer function with threshold level, a fixed-size, fixed-duration, fixed-form spike, and an absolute refractory period. Inputs Iij (i=1...NI) synapses for j-th neuron are weighted with Wij (i=1...NI), where Wij 0 is defined for a inhibitory synapse, Wij = 0 for an inactive or non-existent synapse and Wij 0 for an ex- citatory synapse. Outputs Okj (k=1...NO) can have axonic propagation delays Dkj (a delay can be equal to zero). Firing of neurons in a network follows diffusion process, according to propagation delays; random fluctuations in several processes can be simulated. Several learning algorithms can be implemented explicitly with NEUROL; a Hebbian synapse-strength reinforcement rule has specific language support now. NEUROED's basic specifications are: a) written in Turbo BASIC 1.0 for IBM-PC compatible machines, with CGA monochrome graphics display and optional numerical coprocessor; b) capacity of 100 neurons and 10.000 synapses; c) three neural tissue layers: input, processing and output. d) real-time simulation of neural tissue dynamics, with three display modes: oscilloscope mode (displays membrane potentials along time for several cells simultaneously); map mode (displays bidimensional architecture with individual cells, showing when they fire) and Hinton diagram (displays interconnecting matrix with individual synapses, showing when they fire); e) Realtime, interactive modification of net parameters; and f) capability for building procedures, functions and model libraries, which reside as external disk files. NEUROED and NEUROL are easy to learn and to use, intuitive for neuroscientists, and lend themselves to modeling neural tissue dynamics for teaching purposes. We are currently developing a basic "library" of NEUROED models to teach basic neurophysiology to medical students. Implementations of NEUROED and for parallel hardware are also under way. (Presented at the Fourth Annual Meeting of the Brazilian Federation of Biological Societies, Caxambu, MG, July 1991) ----------------------------------------------------------- A CASCADED NEURAL NETWORK MODEL FOR PROCESSING 2D TOMOGRAPHIC BRAIN IMAGES Dourado SC and Sabbatini RME Center for Biomedical Informatics, State University of Campinas, P.O. Box 6005, 13081 Campinas, SÆo Paulo, Brazil. Artificial neural networks (ANN) have demonstrated many advantages and capabilities in applications involving the processing of biomedical images and signals. Particularly in the field of medical image processing, ANNs have been used in several ways, such as in image filtering, scatter correction, edge detection, segmentation, pattern and texture classification, image reconstruction and alignment, etc. The adaptive nature of ANNs (i.e., they are capable of learning) and the possibility of implementing its function using truly massive parallel processors and neural integrated circuits, in the future; are strong arguments in favor of investigating new architectures, algorithms and applications for ANNs in Medicine. In the present work, we are interested into designing a prototype ANN which could be capable of processing serial sections of the brain, obtained from CT or MRI tomographs. The segmented, outlined images, representing internal brain structures, both normal and abnormal, would then be used as an input to a three-dimensional stereotaxic radiosurgery planning software. The ANN-based algorithm we have devised was initially implemented as a software simulation in a microcomputer (PC 80386, with VGA color graphics and a 80387 mathematical coprocessor). It is structured as a compound ANN, comprised by three cascading sub-networks. The first one receives the original digitized image, and is a one-layer, fully interconnected ANN, with one processing element (PE) per image pixel. The brain image is obtained from a General Electric CT system, with 256 x 256 pixels and 256 gray levels. The first ANN implements a MHF lateral inhibition function, based on a convolution filter of variable dimension (3 x 3 up to 9 x 9 PE's), and it is used to iteratively enhance borders in the image. The PE interconnection (i.e. convolution) function can be defined by the user as a disk file containing a set of synaptic weights, which is read by the program; thus allowing for experimentation with different sets of coefficients and sizes of the convolution window. In this layer, PE's have synaptic weights varying from -1 to 1, and the step function as its transfer function. Usually after 2 to 3 iterations, the borders are completely formed and do not vary any more, but are too thick (i.e., the trace width spans several pixels). In order to thin out the borders, the output of the MHF ANN layer is subsequently fed into a three-layer perceptron, which was trained off-line using the backpropagation algorithm to perform thinning on smaller straight line segments. Finally, the thinned out image obtained pixel-wise at the this ANN's output is fed into a third network, also a three-layer perceptron trained off- line using the backpropagation algorithm to complete small gaps ocurring in the image contours. The final image, also 256 x 256 pixels with 2 levels of gray, is passed to the 3D slice reconstruction program, implemented with conventional, sequential algorithms. A fourth ANN perceptron previously trained by back-propagation to recognize the gray histogram signature of small groups of pixels in the original image (such as bone, liquor, gray and white matter, blood, dense tumor areas, etc.), is used to false-color the entire image according to the classified thematic regions. The cascaded, multilayer ANN thus implemented performs very well in the overall task of obtaining automatically outlined and segmented brain slices, for the purposes of 3D reconstruction and surgical planning. Due to the complexity of algorithms and to the size of the image, the time spent by the computer we use is inordinately large, preventing a practical application. We are now studying the implementation of this ANN paradigm in RISC-based and vector-processing CPUs, as well as the potential applications of neurochip prototyping kits already available in the market. (Presented at the I Latinoamerican Congress on Health Informatics, Habana, Cuba, February 1992) -------------------------------------------------------- COMPUTER SIMULATION OF A QUANTITATIVE MODEL FOR REFLEX EPILEPSY R.M.E. Sabbatini Center of Biomedical Informatics and School of Medicine of the State University of Campinas, Brazil. In the present study we propose a continuous, lumped- parameter, non-linear mathematical model for explaining the quantitatively observed characteristics of a class of experimental reflex epilepsy, namely audiogenic seizures in rodents, and simulate this model with a especially contrived microcomputer program. In a first phase of the study, we have individually stimulated 280 adult Wistar albino rats with a 112 dB white-noise sound source, and recorded the latency, duration and intensity values of the psychomotor components of the audiogenic reaction: after an initial delay one or more circular running phases usually occurs, followed or not by complete tonic-clonic seizures. In the second step, we performed several multivariate statistical analyses of these data, which have revealed many properties of the underlying neural system responsible for the crisis; such as the independence of the running and convulsive phases; and a scale of severity which is correlated to the value of latencies and intensities. Finally, a lumped- parameter model based on a set of differential equations which describes the macro behavior of the interaction of four different populations of excitatory and inhibitory neurons with different time constants and threshold elements has been simulated in a computer, In this model, running waves, which may occur several times before leading or not to the final convulsive phase, are explained by the oscillatory behavior of a controlling neural population, caused by mixed feedback: an early, internal positive feedback which results in the growing of excitation, and a late negative feedback elicited by motor components of the running itself, which causes the oscillation back to inhibition. A second, threshold-triggered population controls the convulsive phase and its subsequent refractory phase. The results of the simulation have been found to explain reasonably well the time course and structural characteristics of the several forms of rodent audiogenic epilepsy and correlates well with the existing knowledge about the neural bases of this phenomenon. (Presented at the Second IBRO/IMIA International Symposium on Mathematical Approaches to Brain Functioning Diagnostics, Prague, Czechoslovakia, September 1990). -------------------------------------------------------- OUTCOME PREDICTION FOR CRITICAL PATIENTS UNDER INTENSIVE CARE, USING BACKPROPAGATION NEURAL NETWORKS P. Felipe Jr., R.M.E. Sabbatini, P.M. Carvalho- J£nior, R.E. Beseggio, and R.G.G. Terzi Center for Biomedical Informatics, State University of Campinas, Campinas SP 13081-970 Brazil Several scores have been designed to estimate death probability for patients admitted to Intensive Care Units, such as the APACHE and MPM systems, which are based on regression analysis. In the present work, we have studied the potential of a model of artificial neural network, the three-layer perceptron with backpropagation learning rule, to perform this task. Training and testing data were derived from a Brazilian database which was previously used for calculating APACHE scores. The neural networks were trained with physiological, clinical and pathological data (30 variables, such as worst pCO2, coma level, arterial pressure, etc.) based on a sample of more than 300 patients, whose outcome was known. All networks were able to reach convergence with a small global prediction error. Maximum percentages of 75% correct predictions in the test dataset and 99.6 % in the training dataset, were achieved. Maximum sensitivity and specificity were 60% and 80%, respectively. We conclude that the neural network approach has worked well for outcome prognosis in a highly "noisy" dataset, with a similar, if slightly lower performance than APACHE II, but with the advantage of deriving its parameters from a regional dataset instead from an universal model. The paper will be presented at the MEDINFO'92 workshop on "Applications of Connectionist System in Biomedicine", September 8, 1992, in Geneva, Switzerland. ============================================================== Reprints/Preprints are available Renato M.E. Sabbatini, PhD Center for Biomedical Informatics State University of Campinas SABBATINI@CCVAX.UNICAMP.BR SABBATINI@BRUC.BITNET