Development of artificial neural network based parametric models for biomedical signals and images analysis /
Parametric models, in form of rational system transfer function, have been successfully applied to solve various problems in different fields of human endeavor. In particular, a lot of research work has been carried out on both the autoregressive (AR) and autoregressive moving average (ARMA) models...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
Gombak, Selangor:
Kulliyyah of Engineering, International Islamic University Malaysia,
2010
|
| Subjects: | |
| Online Access: | http://studentrepo.iium.edu.my/handle/123456789/4586 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Parametric models, in form of rational system transfer function, have been successfully applied to solve various problems in different fields of human endeavor. In particular, a lot of research work has been carried out on both the autoregressive (AR) and autoregressive moving average (ARMA) models as they have been successfully used in many areas of applications such as biomedical signals analysis, system identification, geophysical signal processing, speech analysis just to mention a few. ARMA model, by its structure gives a good parsimonious representation of signals, however it is difficult to estimate its parameters. On the other hand, estimation of AR parameters is relatively easier to accomplish but suffers from spectral line splitting, poor resolution and inaccurate frequency estimates, especially for noisy short data length. In this thesis, artificial neural network (ANN) and genetic algorithm (GA) have been proposed for estimating AR and ARMA model parameters. The motivation for this work lies in the ability of ANN to learn complex mappings between input and output signals which is very difficult to encompass in the conventional model parameters estimation techniques. Development of the ANN-based parametric models has been divided into two, namely real-valued neural network (RVNN) and complexvalued neural network (CVNN) and is applicable to real-valued data and complexvalued data respectively. In both cases, model parameters determination has been carried out by a process of learning from input and output data presented to the neural network. A new model order determination technique has also been investigated in this thesis. GA has been utilized in optimizing the model coefficients obtained from the trained network and in the determination of the optimal network structure and model order. The proposed techniques produce improved performance over the existing techniques especially in accurate estimation of the model coefficients at low signal to noise ratio (SNR). In addition, the newly introduced ANN-based AR model determination techniques do not suffer from poor frequency resolution at low SNR and spectral line splitting associated with some of the existing AR model parameters determination techniques. Furthermore, the proposed hybrid RVNN-based ARMA model has been successfully used to resolve closely spaced frequencies at low SNR. It is well known that magnetic resonance images (MRI) reconstruction using inverse discrete Fourier transform (IDFT) often leads to poor resolution, introduction of artifacts and Gibb's phenomena in the reconstructed images. Though CVNN has been used to alleviate some of these shortcomings, the reconstructed images still contain some isolated spikes. In this thesis, the newly proposed CVNN based parametric models have been applied to MRI for data extrapolation and reconstruction so as to recover the truncated high frequency signals and remove isolated spikes in the reconstructed images. The results obtained when the proposed models are applied to MRI images show improved performance as almost 95% of the truncated high frequency signals were recovered. This is a significant achievement as lower values have been reported in the literature. Furthermore, this thesis considers the use of ANN-based AR modeling technique for constructing a plane to separate members of two sets for two types of applications namely PIMA diabetes and liver disorder classification problem. Results obtained from the application of the proposed algorithm on PIMA diabetes dataset give an accuracy of 83.16% as compared to 82.60% in the previous work. Similarly, in the case of liver disorder diagnosis, accuracy of 74.65% has also been obtained in this work. Finally, CVNN-based CAR model and RVNN-based AR models have been applied to complex boundary and centroidal signatures for shape classification and mould infested images analysis. Hybrid techniques involving Fourier descriptors and RVNN-based ARMA model or CVNN-based CARMA model for dynamic mould growth analysis and prediction have also been discussed in this thesis. Application of structural similarity measures on the reconstructed mould infested image using the proposed models is found to be more computationally efficient as relatively lower number of model coefficients have been used in the analysis, characterization and reconstruction of the original spatial domain boundary pixels. |
|---|---|
| Item Description: | Abstracts in English and Arabic. "A thesis submitted in fulfilment of the requirement for the degree of Doctor of Philosophy (Engineering)"--On t.p. |
| Physical Description: | xxiv, 267 leaves : ill. (some col.) ; 30 cm. |
| Bibliography: | Includes bibliographical references (leaves 230-243). |