Elsevier

Applied Soft Computing

Volume 47, October 2016, Pages 151-167
Applied Soft Computing

A package-SFERCB-“Segmentation, feature extraction, reduction and classification analysis by both SVM and ANN for brain tumors”

https://doi.org/10.1016/j.asoc.2016.05.020Get rights and content

Highlights

  • Brain tumors as segmented regions of interests (SROIs) by content based active contour model (CBAC).

  • Feature extraction—intensity and texture based features.

  • Feature reduction by Genetic Algorithm.

  • Classification by Hybrid Models-GA-SVM and GA-ANN.

Abstract

The objective of this experimentation is to develop an interactive CAD system for assisting radiologists in multiclass brain tumor classification. The study is performed on a diversified dataset of 428 post contrast T1-weighted MR images of 55 patients and publically available dataset of 260 post contrast T1-weighted MR images of 10 patients. The first dataset includes primary brain tumors such as Astrocytoma (AS), Glioblastoma Multiforme (GBM), childhood tumor-Medulloblastoma (MED) and Meningioma (MEN), along with secondary tumor-Metastatic (MET). The second dataset consists of Astrocytoma (AS), Low Grade Glioma (LGL) and Meningioma (MEN). The tumor regions are marked by content based active contour (CBAC) model. The regions are than saved as segmented regions of interest (SROIs). 71 intensity and texture feature set is extracted from these SROIs. The features are specifically selected based on the pathological details of brain tumors provided by the radiologist. Genetic Algorithm (GA) selects the set of optimal features from this input set. Two hybrid machine learning models are implemented using GA with support vector machine (SVM) and artificial neural network (ANN) (GA-SVM and GA-ANN) and are tested on two different datasets. GA-SVM is proposed for finding preliminary probability in identifying tumor class and GA-ANN is used for confirmation of accuracy. Test results of the first dataset show that the GA optimization technique has enhanced the overall accuracy of SVM from 79.3% to 91.7% and of ANN from 75.6% to 94.9%. Individual class accuracies delivered by GA-SVM are: AS-89.8%, GBM-83.3%, MED-95.6%, MEN-91.8%, and MET-97.1%. Individual class accuracies delivered by GA-ANN classifier are: AS-96.6%, GBM-86.6%, MED-93.3%, MEN-96%, MET-100%. Similar results are obtained for the second dataset. The overall accuracy of SVM has increased from 80.8% to 89% and that of ANN has increased from 77.5% to 94.1%. Individual class accuracies delivered by GA-SVM are: AS-85.3%, LGL-88.8%, MEN-93%. Individual class accuracies delivered by GA-ANN classifier are: AS-92.6%, LGL-94.4%, MED-95.3%. It is observed from the experiments that GA-ANN classifier has provided better results than GA-SVM. Further, it is observed that along with providing finer results, GA-SVM provides advantage in speed whereas GA-ANN provides advantage in accuracy. The combined results from both the classifiers will benefit the radiologists in forming a better decision for classifying brain tumors.

Graphical abstract

An interactive computer aided dignostic (CAD) system for assisting inexperience young radiologists is developed. The difficulty in brain tumors classification is due to similar size, shape, location, hetrogeniety, presence of oedema, cystic and isointense regions has been the key feature of this research. Genetic Algorithm is employed as it is an easy concept and is well understood by radiologists without going in much depth of engineering.

  1. Download : Download high-res image (193KB)
  2. Download : Download full-size image

Introduction

Brain tumor classification includes categorization of primary and secondary tumors into different classes. Primary tumors originate in the brain itself like Astrocytoma (AS), Glioblastoma Multiforme (GBM), Meningioma (MEN), Medulloblastoma (MED) etc. Secondary brain tumors or metastases (MET) are the cancer cells that originate from another part of the body and have spread to the brain [1].

Magnetic resonance (MR) images obtained from different excitation sequences like T1, T2, post contrast T1, FLAIR provide texture and intensity information of brain tumors. Of all these sequences, post contrast T1 weighted MR images provide better visualization of brain tumors than the other ones. The post contrast T1 images are obtained after intravenous administration of 0.15–0.20 ml/kg of MR contrast material-(Gadolinium based) in patients. Even though the Gadolinium based contrast material injected is consistent across the patients, the degree of enhancement of tumors is different. Tumors may either show complete solid enhancement or peripheral rim of enhancement or no enhancement as shown in Fig. 1 [2]. A combination of two or more sequences provide better accuracy in delineating and classifying brain tumors; however, these sequences may not always differentiate between the tumor and the associated perilesional oedema. Post contrast T1-weighted images delineate the tumor well along with the demonstration of the extent and character and are thus considered relatively better for brain tumor classification [2], [3].

Radiologists according to their clinical experience classify brain tumors in MR images using pathological details which include signal intensities and texture patterns on MR images [1], [4], [5]. The signal intensities of tumors are defined below:

  • Isointense—The tumors which show isointense property have same intensity as that of brain tissues.

  • Hypointense—The tumors which illustrate hypointense property are darker than the brain tissues.

  • Hyperintense—The tumors which exemplify hyperintense property are brighter than the brain tissues.

The tumor texture may be homogeneous or heterogeneous [4], [5]. These terms are defined below:

  • Homogenous—Tumors show relatively similar signal intensity/brightness in their entire extent.

  • Heterogeneous—Tumors show areas of different signal intensity/brightness (necrotic and cystic part) within themselves.

Pathological details of different tumors on post contrast T1-weighted images are discussed below [1], [4], [5].

  • Astrocytoma—Astrocytoma can be homogenous or heterogeneous lesion. The homogeneous, lesions show a homogeneously similar intensity without significant post contrast enhancement while heterogeneous lesions show variable post contrast enhancement. These lesions especially homogeneous ones are low grade and have good prognosis after surgical resection.

  • Glioblastoma Multiforme—It is a heterogeneous intra-parenchymal mass showing thick peripheral enhancement. There may be associated central necrosis with significant neo-vascularity.

  • Medulloblastoma—It is a childhood tumor, most commonly seen in the posterior fossa. It may be homogenous or heterogeneous lesion in relation to the fourth ventricle with moderate post contrast enhancement.

  • Meningioma—Meningioma is a relatively homogenous well defined extra-axial moderately enhancing mass lesion and shows early enhancement which persists into the delayed phase after intravenous administration of contrast. It compresses and displaces the cortex and the subarachnoid space including its contents i.e., blood vessels.

  • Metastatic tumor—It is a secondary tumor of brain showing thick ring or solid enhancement. It has well defined margins and causes significant perilesional oedema and mass affect on the adjacent normal brain parenchyma.

The basic knowledge of signal intensity patterns is essential as also the principles by which various sequences describe tumor morphology before one proceeds for brain tumor classification. The different types of tumors taken in this study with their signal intensity, texture pattern and enhancement criteria (solid/peripheral/no) on T1-post contrast images are shown in Fig. 1.

The segmentation, feature extraction and classification are the important aspects of any medical decision support system or computer aided diagnostic (CAD) system. Different Computer Aided Techniques (CATs) proposed for marking brain tumor regions on the images, consist of automatic and semi-automatic segmentation methods [6], [7], [8], [9], [10], [11], [12], [13], [14], [15].

Fetcher Heath et al. [6] developed an automatic segmentation method for non enhancing tumors based on fuzzy clustering algorithm. The domain knowledge obtained by training the system and fuzzy clustering were employed for tumor segregation. The segmentation was performed in axial plane of T1, T2 and proton density images. Similar technique was developed by Dou et al. [7] for segmenting Glial cerebral tumors based on fuzzy information collected from different MR sequences. The tumor characteristics, fusion and adjustments were done by employing fuzzy models and operators. However, inputs from different MR sequences were considered.

Yu and Fan [8] partitioned the image into dark, gray and white sub segments by employing fuzzy sets. The image was divided into two groups based on membership functions which lead to the formation of fuzzy sets. The method was able to segment the desired object. Jaffar et al. [9] developed a method of segmenting lung boundaries on CT images. This method was based on fuzzy entropy and morphology. Promising segmentation results were obtained for segregating lungs on CT images. In this technique, optimal threshold was determined by incorporating fuzzy entropy. Similar technique based on knowledge based system was developed by Clark et al. [10]. GBM tumors were segmented based on cluster formation. The cluster centers acted as inputs to a rule-based expert system. Analysis of multispectral histograms separated the tumor region from the normal regions of the brain. However, the time constraint was a major drawback. The above methods took a lot of time for calculating optimal image parameters.

Research based on watershed algorithm segmenting medical images was carried by Grau et al. [11]. The slope information as well as neighboring pixels information was based on the prior knowledge of the pixels. The major limitation of this method is its reliance on prior information.

It is observed that the automatic methods segment objects/tumors which show either solid or peripheral enhancement and are homogeneous in nature. Moreover, these methods are unable to segment heterogeneous objects/tumors with complex boundaries and require high computational time.

In case of semiautomatic segmentation methods, region of interest (ROI) is user defined; therefore, computational time is less for these methods. Few of the researchers which developed semi-automatic segmentation methods for marking object (tumor) boundaries are discussed below.

Semi automatic segmentation method was developed by Xu and Prince [12] based on gradient vector flow (GVF). The GVF was based on geodesic active contours [13]. The GVF contour was bidirectional in nature which prevented the contour from escaping through weak periphery or small boundary gaps. However, the contour was not able to evolve itself at saddle points. This method was tested on MR image of the human heart where it failed to capture the exact boundary due to the presence of irregularity in the periphery. The weak edges acted as false edges for the desired object to be segmented. Wang et al. [14] proposed the curve formation method for capturing the desired boundary based on external force field. This field is generated by the fluid flowing along the object boundary therefore; the method was named as fluid vector flow (FVF). FVF worked well only on high intensity tumor regions, however, it was unable to segment heterogeneous and isointense tumors.

A level set approach was proposed by Xie et al. [15] for segmenting oedema and brain tumors. The curve formation was a function of region and boundary information for providing propagation force and stopping function. However, the tests were performed on tumors which show strong edge information such as Meningiomas and High Grade Gliomas. Another model, Magneto-static active contour model (MAC) based on level set theory was proposed by Xie and Mirmehdi’s [16]. This method was based on the magneto-statics magnetic contacts between the active curve and the object borders. MAC was able to recognize weak edges and irregular boundaries. However, it lead to the formation of many zero level curves as it detected even the weak boundaries during segmentation process.

The general limitations of the above methods are: (i) segmented only homogenous tumors/objects on medical images (ii) failed in case of presence of complex boundaries (iii) detected many false edges leading to multiple small boundaries (iv) did not stop when there were large variations in the gradients along the boundary of the tumor (v) pre-convergence and over-segmentation problems existed in case of irregular tumor/object boundaries.

In this paper, content based active contour model (CBAC) is proposed by Sachdeva et al. [17]. This is a semi-automatic segmentation model based on parametric approach. The content and edge of an image forms the driving criterion of the active contour. Both intensity and texture inside and outside the tumor are considered. Therefore, all the above limitations are dealt with. CBAC has already been tested on 600 images of different tumor types and is discussed briefly in Section 3.1. Segmentation of tumor regions by CBAC forms the integral step before classification of brain tumors.

While interacting with the radiologists, residents and fellows of neuroradiology at Post Graduate Institute of Medical Education and Research (PGIMER) India, it was observed that both the tasks of marking tumor boundaries and further classification are affected by subjective variations which may lead to confliction. Further, ambiguity lies in interpreting tumor class and exact boundary due to similar tumor morphology, heterogeneity, presence of cerebro spinal fluid (CSF). Also, iso/hypo/hyper properties shown by different tumors of the same class lead to difference in opinion. Therefore, an interactive CAD system is developed for inexperienced radiologists & medical students to assist them in both segregating as well as classifying brain tumors.

The proposed CAD system is developed for segmenting and classifying brain tumors on magnetic resonance (MR) images. The system consists of four modules as shown in Fig. 2. In the first module marking of tumor regions by content based active contour (CBAC) model is proposed [18]. The tumor regions are saved as segmented regions of interest (SROIs). The second module consists of feature extraction module in which intensity and texture features are extracted from the SROIs. Thirdly feature selection using Genetic Algorithm (GA) is proposed where GA is used to select a set of salient features from input features. The selected features are used as inputs to support vector machine (SVM) and artificial neural network (ANN). Finally, the classification module consists of classifying brain tumor classes by using SVM and ANN.

The proposed system employs CBAC which segments the tumor boundaries irrespective of heterogeneity, weak & false edges and on any tumor type with a minimum time. CBAC reduces intra and inter observability issues to a large extent. Further, hybrid models namely GA-SVM and GA-ANN are proposed for classification using Genetic Algorithm (GA) with support vector machine (SVM) and artificial neural network (ANN). GA reduces the feature set and combination of GA with SVM and ANN classifies even those tumors which have similarity in shape, location, size, morphology and enhancement. Further, the classification analysis by SVM and ANN depicts that the results can be verified by both the models if still ambiguity persists. The methods are tested on a public domain dataset-Surgically Planning Laboratory (SPL) dataset by Harvard Medical School. Boston, MA, USA [48], [49] and a dataset collected from a Govt. Post Graduate Institute of Medical Education and Research (PGIMER), India. Post contrast T1 MR images are preferred for maintaining uniformity in the intra and inter tumor groups with respect to MR images. Large dataset of 428 images is collected from PGIMER so as to increase the numbers of samples in training and testing phase. Further, the pathological details of the tumors are utilized for extracting the features from Post contrast T1 MR images. The proposed system is steadfast, proficient and provides definitive results. The contentment in results of the developed CAD system is shown by the expert radiologists of PGIMER, Chandigarh, India and Doon MRI Center, Dehradun, India. The CAD system will afford medical professionals in important clinical judgments and for better performance in terms of accuracy and time.

This paper is organized in the following main sections. In Section 2, background theory is discussed. Proposed methodology which includes marking of tumor regions by CBAC and feature extraction techniques is given in Section 3. This section also includes selection using genetic algorithm and the classifier(s) (SVM and ANN) details. Further, standard classifiers are also given in Section 3. Dataset details and software implementation are illustrated in Section 4. In Section 5, experimental set up and in Section 6 results and discussions are given. The paper is concluded in Section 7 and future scope is given in Section 8.

The methodology of proposed CAD system includes segmentation- segregating tumor boundaries, feature extraction- extraction of intensity and texture features, feature reduction- removal redundant features and finally classification- classifying data into different classes based on the ground truth provided by the medical expert.

Section snippets

Background theory

MR images are categorized into various classes such as normal (without tumor or disease) and abnormal (affected from disease-tumor, Alzheimer etc.), or based on tumor tissues such as benign and malignant [18], [19], [20], [21], [22], [23], [24]. The classification is also done on the basis of tumor classes termed as multiclass classification of brain tumors such as Glioma, Astrocytoma, Meningioma, Medulloblastoma etc. [25], [26], [27], [28], [29], [30], [31], [32], [33]. Brain tumor

Proposed method

The proposed system is developed to assist radiologists in classifying brain tumors on MR images as shown in Fig. 2. The system consists of three modules: (i) marking of tumor regions by content based active contour (CBAC) model and saving them as segmented ROIs (SROIs) (ii) feature extraction module from tumor regions (ii) feature selection using Genetic Algorithm (GA) where GA is used to select a set of salient features from input features (iii) classification module using SVM and ANN. The

Dataset and software implementation

The dataset and software details are discussed in the following subsections.

Experimental set up

Two set of experiments are performed to test the classification performance of proposed hybrid GA-SVM and GA-ANN approach over SVM and ANN.

Results and discussions

The results obtained from both the experiments are given in detail in the following subsections.

Conclusion

The methods namely hybrid ‘GA-SVM’ classifier and ‘GA-ANN’ classifier by amalgamating features of GA and SVM, GA and ANN respectively are proposed for multiclass classification of brain tumors. The tumor regions are marked by CBAC. The proposed methods make use of intensity and texture features. The features extracted are based on the pathological details provided by the radiologists. The detailed experimentation is performed for testing purpose on both primary and secondary tumors using

Future scope

This system performs the broad classification of brain tumors. In future, for more crisp classification, dataset comprising of other classes like Gliomas and subclasses like Gliomas Grade II and Gliomas Grade III of tumors will be added to the database and further study will be performed. However, data collection may take a long span of time. A generalized diagnostic system irrespective of data, data type (tumor grades), no. of samples and utilization and testing of other pattern recognition

Acknowledgments

For domain knowledge expertise regular discussions are being held at PGIMER, Chandigarh, India. The authors would like to thank Dr. Harish Bhatia, Doon MRI Centre, Dehradun, India and PGIMER, Chandigarh, India for providing clinical inputs.

References (49)

  • M. Idrissa et al.

    Texture classification using Gabor filters

    Pattern Recognit. Lett.

    (2002)
  • C. Huang et al.

    A GA-based feature selection and parameters optimization for support vector machines

    Expert Syst. Appl.

    (2006)
  • S.K. Warfield et al.

    Adaptive template moderated, spatially varying statistical classification

    Med. Image Anal.

    (2000)
  • P. Kleihues et al.

    The new WHO classification of brain tumours

    Brain Pathol.

    (1993)
  • L. Keijer et al.

    Tissue characterization of intracranial tumors by MR imaging: In vivo evaluation of T1- and T2-relaxation behavior at 1.5 T

    Acta Radiol.

    (1991)
  • L. Kjer et al.

    Texture analysis in quantitative MR imaging: tissue characterisation of normal brain and intracranial tumours at 1.5 T

    Acta Radiol.

    (1995)
  • A.G. Osborn

    Osborn's Brain: Imaging, Pathology, and Anatomy

    (2012)
  • S.W. Atlas

    Magnetic Resonance Imaging of the Brain and Spine

    (2008)
  • L.M. Fetcher Heath et al.

    Automatic segmentation of non-enhancing brain tumors

    Artif. Intell. Med.

    (2001)
  • H.Y. Yu et al.
    (2008)
  • M.A. Jaffar et al.

    Fuzzy entropy and morphology based fully automated segmentation of lungs from CT scan images

    Int. J. Innov. Comput. Inf. Control

    (2009)
  • M.C. Clark et al.

    Automatic tumor segmentation using knowledge-based techniques

    IEEE Trans. Med. Imaging

    (1998)
  • V. Grau et al.

    Improved watershed transform medical image segmentation using prior information

    IEEE Trans. Med. Imaging

    (2004)
  • C. Xu et al.

    Snakes, shapes, and gradient vector flow

    IEEE Trans. Image Process.

    (1998)
  • Cited by (110)

    • Adaptive features selection and EDNN based brain image recognition on the internet of medical things

      2022, Computers and Electrical Engineering
      Citation Excerpt :

      The convolutional neural system hypothesis gets through the bottleneck of manual determination highlights and regularly outperforms the conventional acknowledgement calculation in order and acknowledgement. The practical application in cerebrum picture arrangement and acknowledgement has caused an incredible reaction and new exploration to clinical picture grouping and acknowledgement research imperativeness [18–20]. The remainder of the paper is divided into the following sections: The second section presents a literature review on the presented technique.

    • BrainMNet: a unified neural network architecture for brain image classification

      2024, Network Modeling Analysis in Health Informatics and Bioinformatics
    View all citing articles on Scopus

    The work has been done as a collaborative project to develop an interactive CAD system to assist radiologists under MOU between IIT Roorkee, India and PGIMER, Chandigarh, India.

    View full text