This paper presents a novel brain tumor segmentation method. It is a hybrid of fuzzy c-means clustering algorithm (FCM) and cellular automata model (CA) through the features obtained from gray level co-occurrence matrix (GLCM). Thi s aims to improve the seed growing problem using similarity function generally found in traditional segmentation algorithms. The drawback of traditional similarity function being defined as a distance of pairwise pixels faces the problem of robustness when growing pixels are moving from the seeds. To cope with this problem, fuzzy membership functions obtained by FCM is applied. For performance evaluation, BraTS2013 dataset is empirically experimented throughout in comparisons with the promising compared methods using dice similarity metrics. In this regard, the proposed method shows the outstanding results superior to the compared methods on average. Tumor Detection
INTRODUCTION
Brain tumor segmentation has a purpose to extract the tumor information which leads to treatment and surgical planning. It is a challenging task in clinical practice because the tumors are diverse appearance such as locality, shape and contrast. Currently, multimodal magnetic resonance image (MRI) is widely used for diagnosing the brain tumor pathology due to non-invasive with human and obtained good-soft tissue. It can reveals the clinical information that is appropriate for classifying the tumor tissues. However, brain tumor segmentation still encounters the problems of over and under segmentation due to deformed tumor and ambiguous boundaries. Radiologists necessary require some additional knowledge such as anatomy or pathology for diagnosis correctly. In addition, the emerging of MRI in new patients is a causing of time-consuming for tumor segmentation. Therefore, brain tumor segmentation remains a challenging to develop the algorithm that is more robust and faster.
Numerous segmentation techniques have been proposed in last two decades. A segmentation technique based on cellular automata (CA) is a one of method that effective, fast and simple. Grow-cut is a robust segmentation technique in general image. It is non-promising in medical images because its results obtain a rough boundary. Kim et al. improved the similarity function using a statistical distribution for segmenting brain tumor on MRI. They coped with a rough boundary problem of Grow-cut algorithm. Hamamci et al. proposed the Tumor-cut algorithm for using in a brain tumor segmentation. They established the connection of graph based technique and CA model. It can be solving the shorted path problem on graph cut and random walk algorithms. However, those methods found a problem of wrong segmentation caused from ambiguous boundary of tumor. Because a growing of pixels energy which is examined by similarity value of center pixel and neighboring pixels is decreasing significantly when growing pixels are moving from the initial seeds.
he advantage of FCM algorithm is a motivation to establish the connection of seed segmentation based on CA with fuzzy membership function (see in Fig. 4). The overview of proposed method is shown in Fig. 5. Firstly, image feature is extracted by GLCM that corresponds with the CA based image segmentation. Next step, the image features are clustered by FCM. Subsequently, the similarity function is replaced by membership function. Finally, the image is segmented based on CA rule. A. GLCM Feature GLCM is a feature space which is defined by the frequency of co-occurrence of gray-level. In section 2B introduced a general GLCM for four directions