Elsevier

Applied Soft Computing

Volume 75, February 2019, Pages 21-28
Applied Soft Computing

Modified Genetic Algorithm approaches for classification of abnormal Magnetic Resonance Brain tumour images

https://doi.org/10.1016/j.asoc.2018.10.054Get rights and content

Highlights

  • Three different modified Genetic Algorithm approaches are proposed in this work for feature selection.

  • These approaches are used for Magnetic Resonance brain image classification.

  • Back Propagation neural network is used as the classifier.

  • The proposed approaches are compared with conventional GA approach.

  • An extensive analysis is performed in terms of accuracy, sensitivity and specificity.

Abstract

Genetic Algorithm (GA) is one of the bio-inspired optimization techniques available for practical applications. The increasing necessity for bio-inspired optimization techniques has lead to the development of many innovative optimization techniques. In the backdrop, GA is completely forgotten and rarely used for practical applications. One of the significant reasons for the low preference of GA is the excessive “randomness” associated with this algorithm. The random nature of many processing steps in GA often leads to inaccurate results. The main focus of this research work is to enhance the usage of genetic algorithm for practical applications. Modified GA approaches are used in this work to overcome the drawback of the conventional approaches. In this research work, suitable modifications are made in the existing GA to minimize the random nature of conventional GA. Specifically, the focus of this work is to develop modified reproduction operators which forms the core part of this algorithm. Different binary operations are employed in this work to generate offspring in the process of crossover and mutation process. These binary operations are designed with a specific objective unlike the conventional binary operations in GA which are highly random in nature. The application of these approaches is explored in the context of medical image classification. Abnormal brain images from four different classes are used in this work. The proposed method has yielded 98% accuracy in comparison to other methods. Experimental results show promising results for the proposed approaches in terms of accuracy measures.

Introduction

With the significant increase in the amount of data for computing applications, optimization has become important in all the practical applications. Among the optimization techniques, evolutionary computation-based optimization techniques hold a prominent position due to its numerous advantages. Genetic Algorithm is the first framed evolutionary optimization algorithm which is based on Darwin’s theory of natural evolution. GA has been implemented in various practical applications. However, the success rate of GA is very limited because of the random nature in the mathematical steps of this algorithm. With the advancement in swarm intelligence and other state-of-art evolutionary approaches, the usage of GA for practical applications declined significantly. The main motivation is the lack of ability of GA to guarantee high success rate for practical applications. The low performance of GA is mainly due to the “randomness” associated with many operations in GA. This is the major research gap between GA and other optimization algorithms. Thus, suitable modified GA based approaches are proposed for brain image classification which is the main objective of this work.

However, few research works are being carried out to re-invent the applicability of GA in practical scenarios. Specifically, medical applications are in much need for these improved GA approaches for assisting the computing processes. Automated image-based abnormality detection system involves the concept of feature selection in which GA plays a prominent role. The success rate of the subsequent steps is purely based on the feature selection process. Since conventional GA suffer from few drawbacks, modifications in the existing algorithm is necessary to develop an improved system. Literature survey reveals the availability of few modified GA techniques for biomedical image processing applications. However, the scope for improvement is significantly high in such cases.

Medical image segmentation using a modified Genetic Algorithm is reported in Payel et al. [1]. Prior knowledge about different representations of human organs are combined by GA to perform the segmentation. The authors have applied this method on both Computer Tomography (CT) and Magnetic Resonance (MR) images. However, prior knowledge is not available for most of the applications. Modified GA approach is used for medical image encryption in [2]. The initial population for GA is clustered in this modified approach unlike the conventional GA. This method is applied on MR brain images to validate the results. It may be noted that incorrect clustering can lead to inefficient results. Nagarajan [3] have used a hybrid approach for feature selection in brain tumour images. The concepts of GA and Artificial Bee Colony (ABC) algorithm are used in this work. But, the results are not validated.

Modifications are observed in the fitness function of the proposed approach. Sourav and his research team have developed a hybrid GA approach with the concepts of genetic algorithm and fuzzy logic theory [4]. The number of segments in MR brain image is automatically identified without any prior information in this method.

A hierarchical GA for medical image segmentation is proposed in [5]. Modifications are performed in the structural arrangement of chromosomes which has improved the performance of the brain image segmentation system. A hybrid genetic algorithm for image denoising is reported in [6]. A combination of mathematical methods and genetic algorithms are used in this work for enhancing the quality of the images. Lack of strong quantitative analysis is the major concern of this work. A modified real coded genetic algorithm for segmentation of MR brain images is discussed in [7]. Simulated binary approach-based crossover operations are carried out in this work. An extensive analysis on various bio-inspired methods for medical image processing is reported in [8]. A combination of genetic algorithm and entropy are used to segment MR brain images [9]. However, the method is not tried on noisy images. An attempt to apply genetic algorithm on noisy brain image segmentation applications is made by Ujjwal [10].

A hybrid fuzzy-genetic algorithm-based approach for brain tumour segmentation in MR images is employed in [11]. The proposed approach in this work outperform the conventional GA based approach. Another modified GA based brain image segmentation method is reported in [12]. The number of classes in the fuzzy clustering algorithm is determined using modified genetic algorithm. The combination of region growing and genetic algorithm is used for brain image segmentation in [13]. The initial points of region growing approach is determined using GA in this work. The application of genetic algorithm for image processing in Computer Tomography (CT) images is explored in [14]. Pericardium contour extraction process is performed by genetic algorithm in this work. Modified genetic algorithms are also used for image denoising problems which can be applied on medical images for improving the performance of the overall system [15].

Several researchers have used hybrid genetic algorithms for practical applications. The combination of genetic algorithm and firefly algorithm is used for medical decision support system in [16]. The success rate of this hybrid approach is validated on different types of medical data. The combination of genetic algorithm and learning automata for cancer classification is explored in [17]. Six cancer microarray datasets are used in this work to validate the efficiency of the proposed approach. The combination of Support Vector Machine (SVM) and GA for medical image classification is used in [18]. The proposed approach is analysed in terms of sensitivity and specificity. A survey on the various hybrid evolutionary algorithms is carried out by Drugan [19]. The merits and demerits of the various approaches are discussed in this paper. The combination of discriminant analysis and genetic algorithm is also explored in [20]. Many other related works are available in [21], [22], [23], [24], [25], [26]. These literature works specify the necessity for modified and hybrid GA based approaches.

All the previous works depend only on conventional GA for brain tumour image classification applications. In this work, three modified genetic algorithm-based approaches are proposed for brain image classification. The modified genetic algorithms are used for feature selection in the automated image classification system. The modification in the proposed genetic algorithms are seen in the crossover operators used in this work. In the first method, a combination of binary ‘OR’ and ‘AND’ operations are used in the crossover operator. In the second method, gray code conversion concept is used to generate the offspring. In the third method, segment based binary operation is used in the crossover process. These are the main novel contributions of this work. All three methods are framed with an objective to remove the randomness seen in the conventional GA. The proposed approaches are tested on abnormal brain tumour image classification application. Since brain tumour detection is very crucial for further treatment planning, the success rate of these methods is extremely important. The modified approaches are found to be superior to the conventional GA based approach. Thus, the main contributions of this work are:

  • Three different modified GA approaches are proposed in this work for feature selection.

  • These approaches are used for Magnetic Resonance brain image classification.

  • Back Propagation neural network is used as the classifier

  • The proposed approaches are compared with conventional GA approach.

  • An extensive analysis is performed in terms of accuracy, sensitivity and specificity.

The rest of the paper is organized as follows: Section 2 deals with materials and methods, Section 3 covers the concepts of feature extraction, Section 4 shows the feature selection process, Section 5 deals with classifier, Section 6 illustrates the experimental results and discussions and Section 7 provides the conclusions and future scope of this work.

Section snippets

Materials and methods

The framework of the proposed automated system is shown in Fig. 1. Real time abnormal brain tumour images are used in this work. These images are collected from M/s. Devaki Scan centre, Madurai, India.

450 images from four different abnormal categories such as metastase, glioma, astrocytoma and meningioma are used in this work. These images are grey level images with dimension of 256 × 256. Sample images from these four categories are shown in Fig. 2.

An extensive feature set is extracted from

Feature extraction

Feature extraction is one of the significant steps of any image based automatic classification system. It is used to extract the key characteristic features of images from different classes which aid in accurate classification by the classifier. In this work, fourteen textural features are extracted from the images. These images are extracted from the Gray Level Difference Method (GLDM) of the input image. Features extracted from GLDM are called as higher order statistical features. The

Feature selection

Feature selection is one of the predominant steps in any medical image classification system. All the extracted features do not guarantee high accuracy. The presence of irrelevant features reduces the overall accuracy of the system. This feature selection step eliminates this specific drawback and boost the success rate of the overall system. Bio-inspired optimization techniques are often employed for feature selection in medical images. The primitive bio-inspired evolutionary optimization

Classifier

The optimal features are then used to classify the different types of abnormal brain images. In this work, the conventional Back Propagation Neural (BPN) network is used for the classification process. It is a feedforward neural network with supervised training methodology. A three-layer architecture with input layer, hidden layer and output layer are used. The number of neurons in the input layer is equal to the number of input features and the number of neurons in the output layer is equal to

Experimental results and discussions

The software used for the implementation is MATLAB. The experiments are carried out on an Intel Processor with 4 GB RAM. Initially, the dataset is divided into training images and testing images. Table 2 provides the details about the dataset used.

The performance measures are analysed only for the testing images. The performance measures used in this work are: (a) Classification Accuracy (CA), (b) Sensitivity (Se) and (c) Specificity (Sp). The formulae for estimating these performance measures

Conclusions and future scope

In this work, modified GA based approaches are proposed for MR brain image classification system. Three different GA approaches are proposed in this work for feature selection. BPN is used as a classifier for all the GA based methods. An approximate improvement of 4%–6% accuracy is achieved with the proposed approaches over the conventional GA based approach. In addition to the improved accuracy, the number of features in these proposed features are also minimized. This will lead to the reduced

Acknowledgement

The authors thank M/s. Devaki Scan Centre, Madurai, Tamilnadu, India for the image database and result validation.

References (32)

  • PayelG.

    Incorporating priors for medical image segmentation using a genetic algorithm

    Neurocomputing

    (2016)
  • HosseinN.

    Medical image encryption using a hybrid model of modified genetic algorithm and coupled map lattices

    Opt. Lasers Eng.

    (2018)
  • De Sourav, Bhattacharya Siddhartha, Dutta Paramartha, Automatic magnetic resonance image segmentation by fuzzy...
  • Chih-ChinL. et al.

    A hierarchical evolutionary algorithm for automatic medical image segmentation

    Expert Syst. Appl.

    (2009)
  • JonatasL.P. et al.

    An approach based on hybrid genetic algorithm applied to image denoising problem

    Appl. Soft Comput.

    (2016)
  • MarcinW. et al.

    Bio-inspired methods modelled for respiratory disease detection from medical images

    Swarm Evol. Comput.

    (2018)
  • Cited by (70)

    • Multi-class classification using a new Bayesian method

      2023, Model Assisted Statistics and Applications
    View all citing articles on Scopus
    View full text