By Topic

Artificial Intelligence Methods for Biomedical Data Processing, IEE Colloquium on

Date 26 Apr 1996

Filter Results

Displaying Results 1 - 11 of 11
  • Identification of boundaries in MRI medical images using artificial neural networks

    Page(s): 6/1 - 6/6
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (324 KB)  

    In the area of medical imaging, fully-automatic and robust segmentation techniques would have an enormous beneficial impact on clinical practice and research, by decreasing dramatically the manual effort which must otherwise be devoted to this task. Deployment of conventional image processing techniques has not so far led to a fully-automatic solution, although semi-automatic systems do exist. Since no known, robust segmentation algorithm exists, the ability of neural networks to discover regularities and features in complex data is appealing. Indeed, many preliminary attempts at neural segmentation have been described, although none yet achieves the necessary level of performance for routine application. Southampton General Hospital have a requirement to obtain lung-boundary data within an asthma research project. In connection with this requirement, we have previously reported on work in which multilayer perceptrons (MLPs) are trained using backpropagation to segment the region of the lungs in magnetic resonance images of the thorax. This is achieved by training the network to classify voxels as either boundary (voxels on the boundary between lung interior and surrounding tissue) or non-boundary. In this paper, we present the latest results using this technique. We also show how the generalisation performance of the MLP can be improved using a variety of techniques, including weight pruning algorithms View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Neural network and conventional classifiers to distinguish between first and second heart sounds

    Page(s): 3/1 - 3/6
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (368 KB)  

    A technique to distinguish between the first and second heart sounds without the need for a reference ECG is described. The choice of features for presentation to classifiers is discussed and several types of classifier are introduced. Comparative results for each of the classification techniques are given for data sets obtained from both normal and pathological cases. A misclassification rate of 5.76% is obtained using a neural network classifier whereas conventional classifiers are shown to give a relatively poor performance View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Adaptive non-linear filtering of ECG signals: dynamic neural network approach

    Page(s): 1/1 - 1/6
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (328 KB)  

    This paper describes some aspects of using dynamic neural networks in predictive-type ECG filtering in comparison with adaptive linear filters. Two new algorithms are introduced and their performance is compared with normal and temporal backpropagation algorithms in terms of both the signal-to-noise ratio and the signal quality. The results indicate that using variant step size in learning algorithms improves the signal quality View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Assessing the confidence of classification in artificial neural networks

    Page(s): 4/1 - 4/6
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (288 KB)  

    Artificial neural networks (ANNs) have become very popular for data analysis over the past decade. In particular, feedforward neural network classifiers have, it may be argued, become so popular because they can estimate a posteriori probabilities directly by forming a mapping function from the data space to a probability space. If, however, we are to exploit the undoubted utility of ANNs in safety-critical environments then classification performance in itself is not enough. One of the key requirements of any statistical analysis system is to assess its own confidence in a decision. In the field of medical diagnostics, this requirement is paramount. Part of the problem for any Bayesian classifier is the fact that the posteriors, by definition, sum to unity. This means that a classification is made into one of a closed set of classes. If `rogue' data appears then, even if it fails to conform to the statistics of `genuine' data, it will be classified with apparent confidence into one of the output classes. We must, therefore, monitor the confidence in any classification decision. It is possible to further extend the sophistication of error and confidence estimates for ANNs by incorporation of more complex training and inference (using a full Bayesian methodology, for example). This paper looks at some of the issues involved in estimating errors and confidence limits in feedforward networks, and results are presented on an example of muscle tremor classification View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Fully automatic left ventricular myocardial boundary detection in echocardiographic images: a comparison of two modern methods

    Page(s): 5/1 - 5/6
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (372 KB)  

    The echo images in 2D echocardiography have poor noise characteristics and low spatial and grey-value resolutions. Numerous attempts have been made to develop automated algorithms for quantitative analysis and boundary extraction in these images, but as yet none have been developed adequately to be used clinically. This report presents two modern approaches for automatic extraction of the left ventricular (LV) epicardial and endocardial boundaries from short-axis (SA) echocardiographic data, and compares their performance. Both methods use the radial search algorithm in the extraction process. In the AMRBDS (automatic multiresolution boundary detection system), the first stage uses fuzzy logic and the spatial and intensity information of the input image to estimate the LV centre point (LVCP). Then, a novel multiresolution edge detection technique based on the wavelet transform is applied to each one of the radial intensity profiles to extract the most probable and unique LV edge points along them. Median post-filtering and cubic B-spline techniques are employed to produce the final LV boundaries. In the AANNBDS (automatic artificial neural networks boundary detection system), an MLP (multilayer perceptron) is used to detect the most appropriate centre point of the LV. A second MLP is trained to classify each pixel on the radial lines as an inner, outer or non-edge point. Finally, knowledge guided snakes are employed to extract the LV borders by minimization of the snakes' energy function View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • EEG-based assessment of anaesthetic depth using neural networks

    Page(s): 10/1 - 10/6
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (360 KB)  

    To determine the appropriate dose of anaesthetic drug to be used during surgery is a far from trivial problem. Signs of autonomic activity are not very accurate; hence, an alternative technique is required to monitor the depth of anaesthesia. In common with many biological systems, brain activity is a dynamical system having irregular and unpredictable characteristics. The mathematical analysis of the behaviour of dynamical systems has expanded rapidly with the advent of fast computers. This has given rise to a wide range of medical applications, including mathematical analysis of the human EEG. It is in this last genre that the automatic assessment of the depth of anaesthesia, using ongoing EEG as a measure of brain activity, is discussed. One of the most important problems in EEG analysis is the extraction of appropriate features to describe the ongoing signal, and this can be tackled in various ways. The feature extraction stage of the work described in this paper was performed using methods of dynamic systems analysis and involved the extraction of stochastic and dynamic complexity features of the signal. Feature-spaces formed using these two methods were used as input to a radial basis function pattern classifier. We show that, despite the agent specificity of EEG changes in anaesthesia, a useful anaesthetic depth monitor may be created. We present results for two different anaesthetic agents: desflurane and propofol View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A hybrid neuro-fuzzy system for the classification of normal, fusion and PVC cardiac beats in the MIT-BIH database

    Page(s): 2/1 - 2/6
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (364 KB)  

    Hybrid architectures for intelligent systems is as a new direction in the artificial intelligence research which aims at the development of the next generation of intelligent systems. Current research interests in this field focus on integrating neural networks and fuzzy logic to better exploit their strengths as well as expand their application domain. This paper presents a hybrid neuro-fuzzy system (HNFS) for the classification of normal, fusion and PVC (premature ventricular contraction) cardiac beats of patient 208 in the MIT-BIH Arrhythmia Database. The HFNS has a hierarchical topology consisting of three kinds of building blocks-fuzzy neural networks (FNNs), neural networks (NN) and fuzzy systems (FS). The FNNs and NNs are based on the feedforward backpropagation model. The FS is based on classical methods that measure the QRS area, height and RR interval. The first level of the HFNS accomplishes the task of classification of QRS complexes from leads MLII and VI into different classes. In case of the classification output being ambiguous, the QRS pattern is passed to the second HFNS level for final decision-making. A small inference system is developed to support the decision when the classification obtained from MLII and VI differs. Different configurations of FNNs, NNs and FS for both the first and the second level have been examined and tested. Classification results based on sensitivity and predictivity rates are presented and compared to previous approaches View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A comparison of neural network and traditional signal processing techniques in the classification of EMG signals

    Page(s): 8/1 - 8/5
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (356 KB)  

    The decomposition of an electromyographic (EMG) signal is the separation of the complex signal into its constituent motor units' action potentials (MUAPs). The goal of this work is to achieve real-time interpretation of MUAPs. Traditional approaches are relatively successful but suffer from the drawback of being relatively slow and often requiring human intervention. Our work attempts to overcome these problems in two stages. (1) Use traditional methods where appropriate, but speed up the processing. We are currently developing a multiprocessor array using an Analogue Devices ADSP-21060 SHARC processor. (2) Use AI techniques to reduce the amount of human intervention. Work has started on using neuro-fuzzy methods to resolve MUAPs and thus to classify the waveforms for each train. It is likely that the solution to the decomposition problem will require both AI and signal processing techniques. Because of the non-stationary nature of the EMG signal, the neural net requires the weights to be continually modified to reflect the changes in wave shape. All of the methods described in this paper require a certain amount of pre-processing using digital filters and some used simple compression techniques, which if to be realised in real time, will require fast DSP hardware. The pattern recognition aspect of the decomposition has been realised using both traditional methods and AI, but, because of their generalisation abilities, neural nets appear to be a slightly better option View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Intelligent enhancement and interpretation of EEG signals

    Page(s): 11/1 - 11/7
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (352 KB)  

    Describes the development of a technique which identifies frequency clusters within an EEG and calculates for each cluster the following features: amount, organisation, frequency, amplitude, location, symmetry and changes on eye opening. Previously developed techniques to interpret the EEG represent expertise using conventional knowledge-based systems and algorithms. These systems produce discontinuous outputs, switching from one deduction to another when the inputs cross rule boundaries. The technique described in this paper provides more accurate approximation to the expertise by basing the representation of the knowledge and inference on fuzzy sets and fuzzy logic. In such systems, boundaries need not exist. Only at the final stage-linguistic approximation-do any effects of boundaries have any effect, and these can be minimised by appropriate selection of primary fuzzy sets, hedges and connectives. Previous techniques neither take into account the effect of artefacts or adequately model the expertise, which is widely recognised as largely subjective. This paper brings together work carried out in both these areas to produce a system which should provide value to the clinical workplace. The system proposed eliminates the bias in the output by omitting from the clustering procedure frequency peaks which are suspected of having artefact origin by incorporating work from Wu et al. (1994) View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Detection of brain conditions from evoked responses using artificial neural networks

    Page(s): 9/1 - 9/8
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (488 KB)  

    The contingent negative variation (CNV) is a cognitive event-related potential elicited by the presentation of a warning stimulus followed by an imperative stimulus. Neural nets trained on CNV data offer an additional tool for the diagnosis of Huntington's Disease, Parkinson's Disease and schizophrenia, and may also offer a means of detecting and monitoring the pre-onset of Huntington's Disease. In future, it may be possible to quantify and monitor symptoms and to sub-classify the conditions, thus leading to improved drug treatment and patient care. Application of the fuzzy ARTMAP is most promising in this domain View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Analysis of human muscle activity

    Page(s): 7/1 - 7/6
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (332 KB)  

    Illustrates the use of visualisation techniques in the analysis of recordings of muscle activity. Surface electromyogram (EMG) and mechanomyogram (MMG) signals were recorded simultaneously, along with muscle force from the quadriceps of 10 able-bodied subjects during isometric exercise. The MMG signal records muscle vibrations via a small accelerometer attached to the skin surface over the muscle belly and the EMG is recorded using standard electrodes attached to the skin over the muscle. The signals were recorded over four seconds for each of 11 levels of muscle force ranging from 0 to 100% maximum voluntary contraction. Model order estimation methods provide us with an insight into the number of processes involved in the generation of complicated signals. They also provide a means of assessing the suitable dimensionality of input data for subsequent classification methods. The visualisation method discussed in this paper is based on unsupervised learning, i.e. the target labels of the input data are not used during training. Although they provide an indication of the likely success of classification methods, it should be commented that the reduction from a high dimensional space to a lower dimension will generally incur a loss of information. Therefore, classification techniques which make full use of the higher dimensional space, such as radial basis functions, generally give better results than those based purely on the output positions of the Kohonen map View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.