An efficient brain tumor segmentation based on cellular automata and improved tumor-cut algorithm

• The novel gray-level co-occurrence matrix based cellular automata (GLCM-CA) for image transformation was proposed.

• We proposed Improved Tumor-Cut algorithm (ITC) to achieve the higher performance.

• State-of-the-art ITC and GLCM-CA were used for segmentation and evaluation.

• Dice quantitative evaluation metric was implemented on BRaTS2013 training and testing datasets.

Abstract

Over the last few decades, segmentation applied to numerous applications using medical images have rapidly been increased, especially for the big data of magnetic resonance (MR) images. Brain tumor segmentation on MR images is a challenging task in clinical analysis for surgical and treatment planning. Numerous brain tumor segmentation algorithms have been proposed. However, they have still faced the problems of over and under segmentation according to characteristics of ambiguous tumor boundaries. Improving segmentation method is still a challenging research. This paper presents a framework of two paradigms to improve the brain tumor segmentation; image transformation and segmentation algorithm. To cope with ambiguous tumor boundaries, the proposed novel gray-level co-occurrence matrix based cellular automata (GLCM-CA) is presented. GLCM-CA aims to transform an original MR image to the target featured image. It enhances features of the tumor similar to the background areas prior to segmentation. For segmentation, the efficient Tumor-Cut algorithm is improved. Tumor-Cut is an efficient algorithm in tumor segmentation, but faces the problem of robustness in seed growing leading to under segmentation. To cope with this problem, the novel patch weighted distance is proposed in the proposed Improved Tumor-Cut (ITC). ITC significantly enhances the robustness of seed growing. For performance evaluation, BraTS2013 benchmark dataset is empirically experimented throughout in comparison with the state-of-the-art methods using dice quantitative evaluation metrics. Experiments are carried out on 55 real MR images consisting of training and testing datasets. In this regard, the proposed method based on GLCM-CA feature space and ITC provides the outstanding result superior to the state-of-the-art compared methods.

Introduction

Magnetic resonance image (MRI) provides a good soft tissue imaging and valuable information in pathology of brain tumor without invasive from radioactivity. According to diverse characteristics of brain tumor, MRI could reveal a tumor information in various tissues. As shown in Fig. 1, contrast- enhanced T1-weighted (T1C) highlights the tumor core including active tumor and necrotic, whereas T2-weight visualizes the edema tissue in hyper-intense signal. In clinical practice, MRI is more attentions for diagnosis and prognosis of brain disease (Lawler & Chiocca, 2013).

Brain tumor segmentation is an important and challenging task in surgical and treatment planning. In practice, manual segmentation is widely used in clinical routines but it has still an error and time-consuming. In addition, the radiologists have to have some additional knowledge such as pathology and anatomy (Liu et al., 2014). As such, several semi and fully automated segmentation techniques dealing with various variation of tumor have been proposed (Gordillo et al., 2013, Liu et al., 2014; El Sayed et al., 2014). However, there were a few works dealing with the unclear and irregular boundary (Gordillo et al., 2013, Liu et al., 2014, Sachdeva et al., 2012). Moreover, most of the existing algorithms are not flexible for realistic clinical practice (Njeh, Sallemi, Ayed, Chtourou, Lehericy, & Galanaud, 2015) due to non-reliable model building with ill-training data and inflexible parameters in different datasets. In addition, tumor segmentation result reported to date is 70–85% on average on dice metric (Hamamci et al., 2012, Menze et al., 2015). These techniques still provided an over and under segmentation from the ambiguous tumor boundary, which often occurs on edema tissues.

Cellular Automata (CA) are promising model being successfully applied to a variety of advanced researches (Wolfram, 2002, Wongthanavasu and Ponkaew, 2016). More specifically, CA-based segmentation algorithms were introduced to address the global optimization for different datasets which are robust to noise (Huang et al., 2014, Menze et al., 2015). These algorithms establish the connecting of graph-based seed segmentation to CA model. Grow-Cut is a beginning CA-based segmentation proposed by Vezhnevets et al. (2005). It is an efficient algorithm, but has some drawbacks when cope with specific structures which anatomical structure exhibits relative smooth boundaries. Subsequently, Tumor-Cut (Hamamci et al., 2012), which is a CA-based direct-purposed tumor segmentation algorithm, was proposed. It modified the local transition function of Grow-Cut to eliminate the drawbacks. In addition, it figured out the shortest path problem encountered in the graph-based seed segmentation. Tumor-Cut provides an outstanding result superior to other graph-based seed segmentation techniques; Grow-Cut, Graph-Cut and random walker (Hamamci et al., 2012). Unlike other semi-automatic segmentation algorithms, Tumor-Cut requires a minimal degree of user interaction to draw a line over diameter tumor (Huang et al., 2014). Although Tumor-Cut provides an outstanding result for tumor segmentation, it has still faced the problem of robustness in seed growing. CA's state energy defined by the similarity function in Tumor-Cut will decrease significantly when moving from seed pixels. This leads to under segmentation.

Besides, there are attempts to enhance tumor boundaries on an tranforming image. Saad, Abu-Bakar, Muda, Mokji, and Salahuddin (2011) used GLCM feature similar to 3D histogram to segment the brain lesions. They learned and identified the characteristic of abnormal areas, such as tumor, infraction, abscess and haemorrhage, using contour plot based on GLCM feature space. The brain lesions are then segmented by threshold-based technique. Although the method can segment the brain lesions on both hyper-intense and hypo-intense signal, it still provided wrong results from the drawback of threshold-based technique. In addition, Sachdeva et al. (2012) proposed the method to transform MRI to content-based image. They used features from a gray-level co-occurrence matrix (GLCM) to analyze a characteristic of tumor and non-tumor. The predominant texture obtained by GLCM features and original intensity (MRI) values are identified the tumor region. Then, the content-based image was segmented using Active Contour Model (ACM) (Kass et al., 1988, Wang et al., 2009a, Xu and Prince, 1998) for core tumor only. However, the result was in problem with unclear boundary. In additon, Sompong and Wongthanavasu (2014) proposed an image transformation based on GLCM feature prior to segmentation. The proposed local transition function is inferior to the problem that the tumor areas are similar to the background area. The resulting transforming image is segmented in comparison with Tumor-cut algorithm (Hamamci et al., 2012) and active contour model (Wang, Cheng, & Basu, 2009) on small dataset. The work needs to be improved.

This paper presents a novel brain tumor segmentation framework. It consists of two main contributions: (i) MRI enhancement using the proposed GLCM-based CA, and (ii) improvement of the seed growing robustness in Tumor-Cut algorithm. For the first issue, GLCM-based CA (GLCM-CA) is proposed. Firstly, an original MRI is carried out transforming to new features using the proposed GLCM. The tumor characteristic is now learned by GLCM feature space. Secondly, the proposed CA algorithm maps these features to the target image. For the second issue, the improved Tumor-Cut (ITC) is proposed to cope with the robustness of seed growing. In this regard, a novel patch weighted feature is proposed in the similarity function. Subsequently, the traditional distance in Tumor-Cut algorithm is replaced by the proposed patch weighted distance. Ultimately, the ITC implements on the GLCM-CA-based target image. For performance evaluation, the public benchmark datasets from BraTS2013 Challenges (2013) are implemented throughout using dice quantitative evaluation metrics in comparison with the state-of-the-art methods summarized in Table 3 and detailed in Section 5.4.

The paper was organized as follows. Following this section, background related to the proposed method is given in Section 2. Section 3 presents the proposed scheme in details and evaluation metrics. In Section 4, benchmark datasets with detailed information are provided. Section 5 gives extensive experimental results of the proposed and compared methods. Ultimately, conclusions and discussions were given in Section 6.