Muhammad Khusairi Osman

Intelligent medical disease diagnosis using improved hybrid genetic algorithm--multilayer perceptron network.

An improved genetic algorithm procedure is introduced in this work based on the theory of the most highly fit parents (both male and female) are most likely to produce healthiest offspring. It avoids the destruction of near optimal information and promotes further search around the potential region by encouraging the exchange of highly important information among the fittest solution. A novel crossover technique called Segmented Multi-chromosome Crossover is also introduced. It maintains the information contained in gene segments and allows offspring to inherit information from multiple parent chromosomes. The improved GA is applied for the automatic and simultaneous parameter optimization and feature selection of multi-layer perceptron network in medical disease diagnosis. Compared to the previous works, the average accuracy of the proposed algorithm is the best among all algorithms for diabetes and heart dataset, and the second best for cancer dataset.

Detection of mycobacterium tuberculosis in Ziehl-Neelsen stained tissue images using Zernike moments and hybrid multilayered perceptron network

Conventional clinical diagnosis of tuberculosis disease such as manual screening by microbiologist are tedious, laborious and time consuming. Therefore, more research has been carried out to develop technologies that able to automate the detection process. This paper presents an automated approach to tuberculosis bacilli detection in tissue section. The proposed approach employs image processing technique and neural network for the segmentation and detection of tuberculosis bacilli. First, images of tuberculosis bacilli in tissue samples are captured using light microscope after stained with Ziehl-Neelsen staining method. Then colour image segmentation using moving k-mean clustering is used to extract tuberculosis bacilli from the tissue image. Two colour spaces, RGB and C-Y colour, were utilised in order to improve the quality of segmentation and robust against various staining condition. Next, geometrical features of Zernike moments are calculated. From these features, the best features that could detect tuberculosis bacilli with higher accuracy were selected using hybrid multilayered perceptron network. Experimental results demonstrate that the proposed method is efficient and accurate to detect the tubercle bacilli in tissue.

Weather Forecasting Using Photovoltaic System and Neural Network

This paper presents the applicability of Artificial Neural Network (ANN) for weather forecasting using a Photovoltaic system. The main objective is to predict daily weather conditions based on various measured parameters gained from the PV system. In this work, Multiple Multilayer Perceptron (MMLP) network with majority voting technique was used and trained using Levenberg Marquardt (LM) algorithm. Voting technique is widely used in many applications to solve real world problem. Different techniques of voting are used such as majority rules, decision making, consensus democracy, consensus government and supermajority. The way of the voting technique is different depending on the problem involved. Majority voting technique was applied in the study so that the performance of MMLP can be approved as compared to single MLP network. The proposed work has been used to classify four weather conditions; rain, cloudy, dry day and storm. The system can be used to represent a warning system for likely adverse conditions. Experimental results demonstrate that the applied technique gives better performance than the conventional ANN concept of choosing an MLP with least number of hidden neurons.

Tuberculosis bacilli detection in Ziehl-Neelsen-stained tissue using affine moment invariants and Extreme Learning Machine

This paper describes an approach to automate the detection and classification of tuberculosis (TB) bacilli in tissue section using image processing technique and feedforward neural network trained by Extreme Learning Machine. It aims to assist pathologists in TB diagnosis and give an alternative to the conventional manual screening process, which is time-consuming and labour-intensive. Images are captured from Ziehl-Neelsen (ZN) stained tissue slides using light microscope as it is commonly used approach for diagnosis of TB. Then colour image segmentation is used to locate the regions correspond to the bacilli. After that, affine moment invariants are extracted to represent the segmented regions. These features are invariant under rotation, scale and translation, thus useful to represent the bacilli. Finally, a single layer feedforward neural network (SLFNN) trained by Extreme Learning Machine (ELM) is used to detect and classify the features into three classes: ‘TB’, ‘overlapped TB’ and ‘non-TB’. The results indicate that the ELM gives acceptable classification performance with shorter training period compared to the standard backpropagation training algorithms.

A genetic algorithm-neural network approach for Mycobacterium tuberculosis detection in Ziehl-Neelsen stained tissue slide images

This paper describes a method using image processing and genetic algorithm-neural network (GA-NN) for automated Mycobacterium tuberculosis detection in tissues. The proposed method can be used to assist pathologists in tuberculosis (TB) diagnosis from tissue sections and replace the conventional manual screening process, which is time-consuming and labour-intensive. The approach consists of image segmentation, feature extraction and identification. It uses Ziehl-Neelsen stained tissue slides images which are acquired using a digital camera attached to a light microscope for diagnosis. To separate the tubercle bacilli from its background, moving k-mean clustering that uses C-Y colour information is applied. Then, seven Hus moment invariants are extracted as features to represent the bacilli. Finally, based on the input features, a GA-NN approach is used to classify into two classes: ‘true TB’ and ‘possible TB’. In this study, genetic algorithm (GA) is applied to select significant input features for neural network (NN). Experimental results demonstrated that the GA-NN approach able to produce better performance with fewer input features compared to the standard NN approach.

Colour Image Segmentation of Tuberculosis Bacilli in Ziehl-Neelsen-Stained Tissue Images Using Moving K-Mean Clustering Procedure

Segmentation of tuberculosis bacilli in Zeihl-Neelsen tissue slide images is a crucial step in computer-assisted tuberculosis bacilli detection. In this paper, an automatic colour image segmentation using moving k-mean clustering was proposed. First, initial filter is used to remove the tissues images which remain blue after counterstaining process. After that, moving k-mean clustering using green component of RGB colour model and R_y component of C-Y colour model are used to segment the TB bacilli from its undesirable background which also remains red even after decolourization process. Then a 5×5 median filter and region growing was used to eliminate small regions and noises. The proposed methods have been analysed for several TB slide images under various conditions. Experimental results indicate that the proposed techniques were successfully segment TB bacilli from its background.

Improved Classification Performance for Multiple Multilayer Perceptron (MMLP) Network Using Voting Technique

This project investigates the capability of multiple multilayer perceptron (MMLP) system with majority voting technique. It is a system which consists of all the best-performed MLPs and a single final output from these MLPs is selected by the voting system. The MLP networks are trained using two types of learning algorithm, which are the Levenberg Marquardt and the Resilient Back Propagation algorithms. The performance of the MMLP networks are calculated based on the percentage of correct classificition. Data from three case studies, triangular waveform classification, breast cancer detection and transport classification, have been used to test the performance of the developed system. The results show that the MMLP system with voting technique has the capability of improving the classification correctness. The most well-known artificial neural network (ANN) architecture is a Multilayer Perceptron (MLP) network which is widely used for solving problems related to data classifications. By approaching these innovated MMLP system with automatic voting, the better classification result will be produced.

Performance comparison of gradient descent and Genetic Algorithm based Artificial Neural Networks training

One of the major issues concerning the Artificial Neural Networks (ANNs) design is a proper adjustment of the weights of the network. There have been a number of studies comparing the performance of evolutionary and gradient based ANNs learning. But the results of the studies, sometime conflicting to each other although the same and standard dataset development had been used. Motivated by this finding, the main objective of this paper is to make another comparison between the variations of gradient descent and Genetic Algorithm (GA) based ANNs training with special emphasize given on the developed algorithm and comparison methodology. Besides, the effect of the crossover operation on GA training is also being investigated. The comparison is done using cancer and diabetes benchmark dataset. The result shows that the overall classification error percentage of the family of GA is slightly better than those of gradient descent on cancer dataset. On the other hand, gradient descent is much better than GA on diabetes.

Performance comparison of clustering and thresholding algorithms for tuberculosis bacilli segmentation

Image segmentation is a key step in most medical image analysis. However, the process is particularly difficult due to limitation of the imaging equipments and variation in biological system. Therefore, accurate and robust segmentation are important for quantitative assessment of medical images in order to achieve correct clinical diagnosis. This paper studies the performance of clustering and adaptive thresholding algorithms for segmenting the tuberculosis (TB) bacilli in tissue sections. Images are obtained by analyzing the Ziehl-Neelsen (ZN) stained tissue slide and capturing using a digital camera attached to a light microscope. Three clustering algorithms namely k-mean clustering, moving k-mean clustering and fuzzy c-mean clustering, and two adaptive thresholding algorithms, Otsu and iterative thresholding, are evaluated for segmentation of TB bacilli. The saturation component, derived from C-Y colour model is utilised as input to these algorithms as it provides good separation between the TB bacilli and the background. The segmentation results are further compared with the manual-segmentation image. Acceptable segmentation accuracy of up to 99.00% was achieved by using the clustering algorithms and the Otsus thresholding. However, k-mean clustering was chosen as it produced the highest TB segmentation rate.

Compact single hidden layer feedforward network for mycobacterium tuberculosis detection

Advances in imaging technology and artificial intelligence have greatly enhanced the research and development of computer-aided tuberculosis (TB) diagnosis system. The system aims to assist medical technologist and improve the accuracy of clinical diagnosis. A typical architecture of a computer-aided TB diagnosis system consists of image processing, feature extraction and classification. Finding an effective classifier for the system has been regarded as a critical topic, in order to improve the detection performance and avoid making false decision. In this study, the recent method called compact single hidden layer feedforward neural network (C-SLFN) trained by an improved Extreme Learning Machine (ELM) is evaluated for detecting the TB bacilli. Six affine moment invariants are extracted from segmented tissue slide images and fed into the C-SLFN. The network is trained and classified the input patterns into three classes: ‘TB’, ‘overlapped TB’ and ‘non-TB’. Further, the study compares the network performance with a SLFN trained using the standard ELM algorithm. The results obtained from this study suggested that the standard ELM still outperformed the C-SLFN in term of classification accuracy. The standard ELM, however requires a large number of hidden nodes compares to the C-SLFN.