Detection of Brain Tumor in 2D MRI: Implementation and Critical Review of Clustering-based Image Segmentation Methods
Lim Jia Qi*, Norma Alias and Farhana Johar
Department of Mathematical Science, Faculty of Science, UTM, Johor Bahru, Malaysia
*Address for Correspondence: Lim Jia Qi, Department of Mathematical Science, Faculty of Science, 81310 UTM, Johor Bahru, Malaysia, Tel: +607-553-4416, ORCiD: https://orcid.org/0000-0002-0496-0143; E-mail: [email protected]
Submitted: 02 January 2020; Approved: 20 January 2020; Published: 22 January 2020
Citation this article: Qi LJ, Alias N, Johar F. Detection of Brain Tumor in 2D MRI: Implementation and Critical Review of Clustering-based Image Segmentation Methods. Int J Cancer Cell Biol Res. 2020; 4(1): 001-013.
Copyright: © 2020 Qi LJ, et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited
Keywords: Image segmentation; Brain MR images; Fuzzy c-means; k-means clustering; Gaussian mixture model segmentation.
Download Fulltext PDF
Background: Image segmentation can be defined as segregation or partitioning of images into multiple regions with the same predefined homogeneity criterion. Brain tumor segmentation is important to isolate the specific brain tumor region for volumetric analysis and characterization of brain tumor types/grades.
Method: This paper explores and investigates several unsupervised image segmentation approaches and their viability and performances in delineating tumor region in contrast enhanced T1- weighted brain Magnetic Resonance Imaging (MRI) scans. First and foremost, raw CE T1-weighted brain MR images are downloaded from a free online database. The images are then preprocessed and undergo an important process called skull stripping. Then, image segmentation techniques such as k-means clustering, Gaussian mixture model segmentation and fuzzy c-means are applied to the preprocessed MR images. The image segmentation results are evaluated using several performance measures, such as precision, recall, Tanimoto coefficient and Dice similarity index in reference to ground truth images.
Results: The highest average Dice coefficient is achieved by k-means (0.189) before post-processing and GMM (0.208) after post-processing.
Conclusion: Unsupervised clustering-based brain tumor segmentation based on just image pixel intensity in single-spectral brain MRI without adaptive post-processing algorithm cannot achieve efficient and robust segmentation results.
Introduction
Human brain possess a very complex structure, and is the central part of the nervous system. Being one of the most important organ in the human body, the presence of brain tumors, be it benign or malignant can be devastating. Brain cancer is rather rare, consisting of merely 1.95% of all cancers in Malaysia [1,2]. Nonetheless, brain cancer is highly fatal. As a matter of fact, the five year relative survival rate is only 34.9% according to the report of Central Brain Tumor Registry of the United States (CBTRUS).
In view of the severity of brain tumor and lack of effective treatment, early diagnosis of the disease can be crucial for the patient recovery and treatment planning for the doctor. Advanced imaging protocols, like Computed Tomography (CT), Positron Emission Tomography (PET), and MRI have provided unprecedented high resolution of medical images that make medical image analysis possible. In this paper, MRI will be the modality of choice because : 1) MRI is non-invasive, 2) produces multiple slices of images of the same tissue region with different contrast by applying different image acquisition protocols and parameters [3], 3) high contrast of soft tissues and high spatial resolution. In this paper, only contrast enhanced T1-weighted MR image will be considered as multi-spectral MR images analysis is expensive and increases odds of segmentation errors due to inconsistency and misalignment [4].
One of the most important steps in brain tumor characterization, detection and visualization is to perform image segmentation to localize the tumor region. Image segmentation is an image processing technique in which several mutually exclusive regions of interests are isolated and segregated based on predefined criteria depending on the algorithm used [5]. According to [6], segmentation of brain tumors can be divided into 3 pathways: manual, semi-automatic and fully automatic. Currently, manual segmentation is often the choice, but the drawbacks are that it is time consuming and heavily relies on the expertise and experience of trained anatomists. This is followed by biopsy (extraction of brain tissue) in conventional clinical trial. One of the primary reason that manual segmentation is still popular is that automated segmentation of brain tumor in medical images can be extremely difficult especially when it comes to delineating “abnormal tissue”. Tumor are often non-rigid and complex in shape, vary in size and position and displays huge variability among patients [6]. Furthermore, intensity homogeneity between surrounding normal tissue and brain tumor and partial volume effect can create confusion within the segmentation algorithm and even radiologists. On the other hand, semi-automated approaches require user input and intervention, which may subject to variability in segmentation results, Even though identification and segmentation of brain tumor in MRI is a great challenge, it is very important in clinical practice for cancer treatment planning and oncology research as a whole. For instance, MRI segmentation can utilized in delineating lesions, image-guided intervention and surgical planning [7]. Owing to high demand and paramount importance of precise image segmentation, various segmentation algorithms with different degree of complexity, either semiautomatic or fully automatic has been put forward.
According to survey done by [7], image segmentation of brain MR images can be categorized into: threshold-based, region-based, pixel classification and clustering, atlas-based, model-based and hybrid method. Lately, deep learning has emerged as a popular method in segmentation of brain tumor [8,9]. Thresholding is a simple image segmentation approach, in which specific intensity value(s), called threshold is determined to separate pixels into their desired class. [10] Proposed segmentation of brain tissues into White Matter (WM), Gray Matter (GM), and Cerebrospinal Fluid (CSF) using thresholding and seed region growing technique. Otsu multi-thresholding segmentation with morphological operations on MR images was proposed by [11]. [12] Applied Berkeley Wavelet Transform (BWT) based brain tumor segmentation. An average of 0.82 dice similarity index was reported.
Region growing refers to image segmentation process that extract groups of neighboring connected pixels with similar intensity and fulfil predefined homogeneity criteria [13]. [14] Applied a modified version of the traditional region growing method or called modified region growing method in 3D brain MR images containing tumors. More accurate boundary detection after segmentation was reported. Seeded region growing was utilized for image segmentation on both 2D and 3D brain MR images in the work of [15,16] performed region growing segmentation with automated seed point selection by taking account of tumor region features. Similarity index and overlap fraction of 0.817 was reported.
Pixel classification and clustering is another type of segmentation method. Under this category, each pixel is treated as an instance or observation and they will be grouped/clustered in classes either based on trained classification models or some similarity criterion. [17] Put forward color converted segmentation with k-means clustering technique. The gray level of brain MR images was converted into a color space image and regions in image are labelled by cluster index. [18] Proposed a Generalized Rough Fuzzy C-Means Method (GRFCM) for brain MR segmentation. Each cluster is characterized by 3 automatically determined rough fuzzy regions. Each region is balanced by a weighting parameter and the bias field is modeled by linear combination of orthogonal polynomials. The proposed algorithm is more robust and produce more accurate segmentation compared to rough c-means and hybrid clustering algorithms. A segmentation algorithm consisting of Self-Organizing Map (SOM) coupled with Learning Vector Quantization (LVQ) was proposed by [19]. Multi-spectral MR images from 20 subjects was collected. Average dice similarity indices are 61% for tumor and 77% for edema.
Atlas based method can a powerful segmentation method if template of human brain for a population of interest is present. This database can provide information on the brain anatomy and serve as references while segmenting new images. [20] Performed atlas-based segmentation of brain tumor images. A total of 5 patient dataset was investigated and an average of 0.96 dice similarity index for tumor was obtained. For model-based segmentation technique, a continuous model is constructed through understanding of priori knowledge about the object of interest, such as size, location and orientation. [21] Proposed a deformable model, called the Charged Fluid Model (CFM) for segmentation of medical images, such as CT and MRI scans. Poisson’s equation was used to guide the evolution of the mathematical model. The main idea of the hybrid approach is to integrate different complementary methods to improve segmentation accuracy. [22] Proposed 2 hybrid segmentation techniques, HASA and EHASA. Both techniques are based on symmetry. The proposed techniques can identify and detect multiple tumors. These techniques do not require any user interaction and are fully automatic. One limitation is that it will not give good results if the tumor is present on the symmetry line. [23] Presented a method for 3D medical image segmentation by combining clustering and classification algorithms. TOP-LBP features and gray level co-occurrence matrix are extracted and random forest is used for classification and segmentation of the necrosis, edema, non-enhanced tumor and enhanced tumor. The dice similarity index for tumor obtained was 0.96 ± 0.01.
Recently, deep learning has quickly emerged as state-of-the-art in biomedical image segmentation due to its superior accuracy, sometimes outperforming clinical experts. However, the biggest bottlenecks for the application of deep learning is the lack of annotated and labelled images [24].
To summarize, although there are wide variety of image segmentation algorithms in the literature, brain tumor segmentation remains a challenging task, due to diverse image content, occlusion, image noise and interference [25]. In view of this, the primary motivation of this study is to assess and compare the performances of several clustering based image segmentation approaches like k-means clustering, Gaussian Mixture Model (GMM) segmentation and fuzzy C-means clustering in terms of execution time and segmentation accuracy (precision, recall, Jaccard index, Dice similarity coefficient). Clustering-based segmentation approaches are unsupervised, which means that the image segmentation techniques do not take class information into consideration. Unsupervised methods are receiving narrow research effort due to difficulty in the integration of prior anatomic knowledge and algorithm [6]. This study aims to bridge the gap by investigating the scalability and practicability of these image segmentation techniques in delineating tumor region in large number of real brain MR images. Lastly, we proposed some modifications to the segmentation algorithms to adapt to brain tumor segmentation and simultaneously minimize parameters adjustment and feedback response (human intervention).
Materials and Methods
Figure 1 shows the flow diagram of the overall processes in this research. The details of the implementation of algorithms involved will also be discussed. All the processes in this research are performed in MATLAB 2017a platform using processor Intel(R) Xeon(R) CPU with 32.0 GB RAM.
Data acquisition
A total of 3064 slices of T1-weighted contrast enhanced MR images from 233 patients was downloaded from https://figshare.com/articles/brain_tumor_dataset/1512427. There are three kinds of brain tumor in the MR images downloaded, namely meningioma (708 slices), glioma (1426 slices), and pituitary tumor (930 slices). The brain T1-weighed CE-MRI dataset was acquired from Nanfang Hospital, Guangzhou, China, and General Hospital, Tianjing Medical University, China, from year 2005 to 2010. The images pixel size are of 0.49×0.49 mm2. The slice thickness is 6 mm and the slice gap is 1 mm [26,27]. Furthermore, the brain tumor in the images has been delineated or segmented by radiologists which can be ground truth images in evaluating and validating the subsequent image segmentation performance.
Brain MR images preprocessing
Medical images are often come with poor contrast and corrupted by noises arise from old MRI machine, image transmission and image digitization [16], extracranial tissue and etc. These factors degrade image quality and inevitably affect subsequent segmentation results regardless of methods employed [19]. In view of the above issues, brain MR images preprocessing is a vital step to remove the unwanted interferences present in raw MR images.
There has been numerous image preprocessing methods proposed for the application in MRI in the literature. Some of them includes median filter [28], Gaussian filter, anisotropic filtering [16], contrast enhancement [12], bias field removal [29], histogram matching [30] and skull stripping [31,32], The justification behind preprocessing of images is to increase signal to noise ratio, enhance visual appearance for both human and machine vision, remove unnecessary parts in the background and preserving important edges and borders and so on. Several preprocessing techniques will be applied in this research. Steps of preprocessing of pathological brain MR images is presented in (Figure 2).
First stage of MR images preprocessing is bias field estimation and removal. Bias field is commonly recognized as low frequency smooth varying signals which accounts for intensity non-uniformity or shading in an image. This phenomenon is usually caused by radio frequency coil non-uniformity, eddy currents, patient positioning and etc [33]. Bias field is undesirable as it reduces the high frequency contents such as borders and contours in an image and cause variation in gray level distribution in the same brain tissues. In this research, second order parametric surface fitting method using Levenberg-Marquadt algorithm, in reference to work of [34] was applied to remove bias field.
Median filter and anisotropic diffusion filter are two primary de-noising methods in brain MR images. Anisotropic diffusion filter are computationally more complex and have several parameters to select to control the operation. Thus, median filter was chosen. A 5×5 median filter was applied to eliminate noise and simultaneously retain important borders. Then, histogram stretching or also known as inner cropping was performed on the image. The goal is to increase contrast without modifying shape of the original histogram.
Skull stripping
Essentially, brain MR images contain non-brain tissue, also known as extra-cranial tissues, such as skull, fat, muscle and connective tissues. The existence of these non-cerebral tissue can be a major obstacle for quantitative analysis later on. Thus, elimination of non-brain tissues, or also known as skull stripping is another vital step before the segmentation of brain MR images can be carried out. In fact, skull stripping is the preliminary step before other further image processing such as image registration [35], tissue classification [36], multiple sclerosis analysis [37], and even diagnosis of neurodegenerative disease like Alzheimer’s disease [38].
Published review studies from [31,39] show that different methodologies have its own advantages and limitations. Even though there is extensive research on skull stripping on brain MRI, majority of the methods do not give satisfactory results over a wide range of scan types due to anatomical variation, imaging artifacts, varying contrast properties and so on [40]. In this study, multi-level thresholding [11] followed by morphology based post-processing method [41] was implemented.
The proposed algorithm consists of 3 phases:
1) Four-class thresholding. The three thresholds, k1, k2 and k3 separating 4 classes are found using thresholding method proposed by [42]. Otsu thresholding is one of the binarization algorithms that was popularly employed in separating foreground and background in an image. The threshold selection is based upon the basis of maximizing inter-class variance and minimizing intra-class variance. The pixels that is above k1 and less than or equal to k3 are believed to be brain region and will be retained for next phase.
2) Morphological operations commences with erosion, followed by largest connected component analysis [43], dilation and lastly morphological flood fill operation. The structural elements used is disc shape with radius of 5 units for both erosion as well as dilation operations. Erosion on the binary images is to remove small disconnected regions, while dilation is to smoothen the boundary of object. With the assumption that the brain is the largest connected component in brain MRI [44], connected component processing was performed and the largest connected object (brain) are preserved. Flood fill operation aims at filling small holes in binary image.
3) Mask the binary image with the preprocessed image which produces skull stripped image to obtain grayscale brain image. The overall processes of preprocessing and skull stripping are depicted in (Figure 3).
ROI (tumor region) segmentation
In the analysis of brain MRI, image segmentation is frequently used for visualizing brain anatomical structures, such as WM, GM, CSF, and tumor region. In this research, the emphasis will be put on segmenting brain tumor from other brain structures. This will be especially helpful in studying the shapes, location and structures of the tumor. Based on the review of previous works, 4 clusters are often chosen as the parameter [45,46] although it is believed that the tumor tissues can further divided into active regions, edema and necrotic core [6,47,48]. In four clusters setting, the clusters with highest mean pixel intensity will be extracted. In addition, based on the review on clustering-based techniques in delineating tumor region, the authors also noticed that some works proposed post-processing while others do not. The post-processing found in literature include morphological filtering [46] and largest connected components [49]. In view of this, this study aims to combine these two methods in delineating the final tumor mask. Table 1 shows the values of parameters used in the clustering algorithms.
K-means clustering: Clustering, being an unsupervised learning method, can be applied in image segmentation and k-means clustering is one of the most popular clustering method in brain tumor detection [50,51]. It is simple and can cope with large number of variables [52]. The parameter, k as its name suggests is a user defined parameter that specify the number of clusters, which can be critical factor in final segmentation result. Low k produces small number of groups that segregate gray matter and white matter, whereas high k probably results in over-segmentation (a lot of disconnected regions) that can increase difficulty in post-processing. Therefore, we perform k-means clustering with different k to evaluate the effect of k on the segmentation results.
The pseudo-code of k-means clustering in segmentation of brain tumor from MRI in this research is shown below:
1) Initialize parameter k, maximum number of iterations, mi, and error tolerance, Ɛ.
2) Randomly generate centroid for each k clusters.
While
Calculate the dissimilarity measures (Euclidean distance) between each centroids with other pixels:
Determine the shortest distance of each pixel with centroids.
Assign each pixel to a specific class, corresponds to the nearest centroid.
Calculate J, objective function.
Update the cluster centroids by formulae below:
If
If
Break while loop.
End if
End if
3) end while
m is the number of row of the image, n is the number of column of the image, d is row vector in which each entry represents distance of each data to its centroid.
GMM based segmentation: GMM is another common probabilistic parametric model in brain tumor segmentation [9,53,54]. Unlike k-means clustering that assigns each pixel to only one label (hard assignment), each pixel is assigned with different probability corresponding to each class specified (soft assignment). Similar to k-means clustering, the number of clusters is arbitrary.
Let denotes the pixel intensities of images and N be the number of pixels of images. The pseudo-code of GMM based segmentation is shown below:
1) Initialize number of clusters, K and maximum number of iteration, .
2) Initialize mean, of each class as pixel intensities picked randomly; initialize standard deviation, of each class as overall standard deviation; initialize prior probability, of each class as; initialize iteration counter as
3) Run the expectation-maximization algorithm
While
Calculate probability of each pixel based on Gaussian probability density function (pdf) for each class:
Calculate the normalized pdf (posterior probability of a certain pixel belonging to class k):
For each cluster, k, update mean, standard deviation and prior probability:
Calculate J, log-likelihood function.
If i ≥ 2
If
Break while loop.
End if
End if
3) end
Fuzzy C-means: Fuzzy c-means algorithm has been widely applied in image segmentation since it was proposed by [55]. The algorithm works by assigning membership to each data points (pixels) according to distance of the data to each cluster centers (centroids). Soft assignment of membership is based on fuzzy logic which gives fuzzy values to any particular pixels to be lying in either of the clusters.
1) Initialize number of clusters, k and maximum number of iteration, mi.
2) Randomly initialize membership matrix, , where i denotes the sample indices and k is the cluster indices with the condition and Initialize fuzzifier, in reference to work of [56] which suggest the use of value in the interval [1.5,2.5].
Let
while
Calculate the cluster centroids, as follows:
Update by the following formula:
If
If
Break while loop.
End if
End if
End while
Post processing: Post processing is conducted to obtain the exact location of tumor region without the assistance of human experts. First, the cluster corresponding to highest mean pixel intensity will be chosen and region with area less than 30 pixels will be eliminated. Next, there are primarily two main steps: morphological opening followed by largest connected component analysis. Morphological opening is performed with disk structuring element of radius 10 to remove small objects and retain the shape and size of bigger objects; largest connected component is assumed to be the tumor region.
Performance evaluation of each image segmentation approaches: Objective and accurate assessment of different image segmentation algorithms on a specific types of image is crucial in helping researchers to properly adjust parameters in each algorithms, get a better picture of strengths and weaknesses and select the best methods. In this study, we will employ empirical discrepancy methods in evaluating the goodness of segmentation results due to availability of gold standard or ground truth binary tumor mask [57]. Widely accepted metric like execution time, precision, recall [58], Jaccard coefficient, and Dice similarity coefficient are used. Let A and B denote the sets of positive pixels (pixels labelled as tumor) by the ground truth and image segmentation methods respectively. n (A) denotes the number of elements present in set A. The details of each performance measures are listed in (Table 2).
Results
In this section, performances of the proposed semi-automatic clustering-based image segmentation algorithms in delineating tumor region will be evaluated based on the ground truth provided alongside the repository. We intend to understand the consistency of the proposed method in large number of 2D single spectral (e.g. in this study, CE T1-weighted MRI) MR images. Section 3.1 discuss the performance of image segmentation without post-processing; section 3.2 discuss the performance of image segmentation with post-processing.
Performance of image segmentation methods without post-processing
As shown in table 3, the average run time of image segmentation methods can be arranged in ascending order such that: k-means, GMM followed by FCM. This demonstrates the superiority of k-means clustering in terms of computational efficiency.
K-means clustering has the highest precision, Jaccard index and Dice coefficient in comparison to GMM and FCM clustering method. Based on the high standard deviation observed in (Table 4-7) and presence of abundant outliers in (Figure 5-8) across all image segmentation approaches, the accuracy in the segmentation results of the proposed methods is inconsistent, suggesting the inability of unsupervised clustering methods in adapting to different MR scans.
Performance of image segmentation methods with post-processing
K-means clustering achieves the highest average precision (0.3), while GMM attains the highest average Jaccard and Dice similarity index (0.173 and 0.208 respectively) according to (Table 8-11). Based on the box-plots of Figure (4-7), at least half of the tumor of MR images are totally missed by the detection algorithms.
Discussion
In summary, the performance of all the clustering-based brain tumor segmentation is highly inconsistent, manifested by the high standard deviation observed in performance measures. This clearly shows that by using image pixel intensity in single spectral MRI alone as features is insufficient in accurately predicting tumor region. Prior knowledge on tumor location and shape should be exploited. It is noted that FCM algorithm suffers from long computational time, has the lowest precision, Dice and Jaccard similarity indices. This finding contradicts to some results reported in literature [54,59, 60]. We would argue this discrepancy arise due to the use of different MRI modality. In our case, contrast enhanced T1-weighted data normally has larger pixel intensity differences between tumor and surrounding tissues (in other words higher contrast to noise ratio [61], while other images might be affected by partial volume effect (homogeneity between tumor and normal tissues), thus fuzzy assigning of pixels to each cluster may be more efficient in that case. The unusual high recall and low precision indicate that FCM tends to overestimate tumor area as shown in (Figure 8).
As can be seen from (Table 12), it can be shown that plenty of previous works regarding the brain tumor segmentation of 2D MR images share these similarities: 1) low number of images, 2) lack of gold standard, thus no empirical results (e.g. Dice, Jaccard) can be reported. It is also noticed that some researches do not include skull stripping and post-processing before and after image segmentation algorithms, which can be critical for accurate and efficient brain tumor segmentation. This study bridges the gap by implementing automatic unsupervised clustering algorithm on large amount (3064 slices) of 2D MR images with ground truth, offering clearer insight on the generalizability of clustering methods in segmentation of brain tumor.
Lastly, the reasons behind the poor performance of the proposed algorithms as compared to what reported in the literature:
1) Lack of shape or intensity priors on the tumors [6,77]. As such, clustering algorithm based solely on pixel intensity is unable detect the exact tumor region that can vary in shapes and location.
2) The unsupervised clustering algorithms proposed in the literature normally exploit anatomic objective measures and include interactive post-processing while our study do not exploit any prior assumptions and implement automatic post-processing.
3) Our method cannot detect necrotic core, normally displayed as hypointense region [65].
4) Skull stripping methods proposed may accidentally remove tumor tissues, especially those that share similar intensity with non-brain tissues as shown in figure 9.
5) The proposed post-processing method is unable to capture ROI accurately across all MR images.
Conclusion
In this paper, we presented several clustering-based brain tumor segmentation using CE T1-weighted MR images. A few common evaluation measures based on manually delineated tumor region are utilized to validate the segmentation performances of each method. Results show that the introduced k-means and GMM is more accurate before and after post-processing. The average dice similarity of k-means (before post-processing: 0.189; after post-processing: 0.2), while for GMM (before post-processing: 0.179; after post-processing: 0.208). Additionally, the segmentation accuracy of this image segmentation approaches can vary greatly among all the 3064 datasets analyzed, based on the high standard deviation in performance metrics. All in all, unsupervised clustering-based brain tumor segmentation with the use of solely image pixel intensity in single-spectral brain MRI without adaptive post-processing algorithm cannot achieve efficient and robust segmentation results.
In view of these limitations of the proposed methods, future research should be oriented towards utilizing more image features like using co-registered multi-spectral MRI features and spatial information (super-pixel segmentation). Sophisticated image segmentation methods like deep learning can be another alternative to enhance the brain tumor segmentation accuracy.
Conflict of Interest
A special thanks to Universiti Teknologi Malaysia (UTM) for the opportunity to carry out the research and Ministry of Education (MOE) for financial support. This project was supported by Research University Grant [Vot Number: 19H03] initiated by UTM and MOE.
- Bondy ML, Scheurer ME, Malmer B, Barnholtz Sloan JS, Davis FG, Il'yasova D, et al. Brain tumor epidemiology: consensus from the brain tumor epidemiology consortium. Cancer. 2008; 113: 1953-1968. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/18798534
- CC Lim G, Yahaya H, Lim T. Second Report of the National Cancer Registry, Cancer Incidence in Malaysia. 2003. http://bit.ly/30MBW93
- Mortazavi D, Kouzani A Z, Soltanian Zadeh H. Segmentation of multiple sclerosis lesions in MR images: a review. Neuroradiology. 2012; 54: 299-320. http://bit.ly/37bgzR4
- Nabizadeh N, Kubat M. Brain tumors detection and segmentation in MR images: Gabor wavelet vs. statistical features. Computers & Electrical Engineering. 2015; 45: 286-301. http://bit.ly/36d9a2e
- Vishnuvarthanan G, Rajasekaran M P, Subbaraj P, Vishnuvarthanan A. An unsupervised learning method with a clustering approach for tumor identification and tissue segmentation in magnetic resonance brain images. Applied Soft Computing. 2016; 38; 190-212. http://bit.ly/2Rww934
- Gordillo N, Montseny E, Sobrevilla P. State of the art survey on MRI brain tumor segmentation. Magnetic Resonance Imaging. 2013; 31: 1426-1438. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23790354
- Despotovic I, Goossens B, Philips W. MRI Segmentation of the Human Brain: Challenges, Methods, and Applications. Computational and Mathematical Methods in Medicine. 2015; 2015: 450341. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25945121
- Havaei M, Davy A, Warde Farley D, Biard A, Courville. Brain tumor segmentation with deep neural networks. Med Image Anal. 2017; 35: 18-31. http://bit.ly/30I5u7x
- Zhao L, Jia K. multiscale CNNs for brain tumor segmentation and diagnosis. Computational and Mathematical Methods in Medicine. 2016; 2016: 8356294. https://www.ncbi.nlm.nih.gov/pubmed/27069501
- Shanthi K J, Kumar M S. Skull stripping and automatic segmentation of brain MRI using seed growth and threshold techniques. http://bit.ly/2ujjJDS
- Chaddad A. Automated feature extraction in brain tumor by magnetic resonance imaging using gaussian mixture models. International Journal of Biomedical Imaging. 2015; 11. http://bit.ly/37b9lML
- Nilesh Bhaskarrao Bahadure, Arun Kumar Ray, Har Pal Thethi. Image analysis for MRI based brain tumor detection and feature extraction using biologically inspired BWT and SVM. International Journal of Biomedical Imaging. 2017; 12. http://bit.ly/36fy3KV
- Wong KP. Medical image segmentation: methods and applications in functional imaging. In J. S. Suri, D. L. Wilson, & S. Laxminarayan (Eds.), Handbook of Biomedical Image Analysis: Volume II: Segmentation Models Part B 2005; pp. 111-182. Boston, MA: Springer US.
- Salman Y M. Modified technique for volumetric brain tumor measurements. J Biomedical Science and Engineering. 2009; 2; 16-19. http://bit.ly/2TFxIOX
- Weglinski T, Fabijanska A. Brain tumor segmentation from MRI data sets using region growing approach. http://bit.ly/2tnr4Cd
- Shanthakumar P, Ganeshkumar P. Performance analysis of classifier for brain tumor detection and diagnosis. Computers & Electrical Engineering. 2015; 45; 302-311. http://bit.ly/2tC7tOu
- Juang LH, Wu MN. MRI brain lesion image detection based on color-converted K-means clustering segmentation. Measurement. 2010; 43: 941-949. http://bit.ly/2RaLx6n
- Ji Z, Sun Q, Xia Y, Chen Q, Xia D, Feng D. Generalized rough fuzzy c-means algorithm for brain MR image segmentation. Comput Methods Programs Biomed. 2012; 108: 644-655. http://bit.ly/2G3Iy9i
- Demirhan A, Toru M, Guler I. Segmentation of tumor and edema along with healthy tissues of brain using wavelets and neural networks. IEEE J Biomed Health Inform. 2015; 19: 1451-1458. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25265636
- Bauer S, Seiler C, Bardyn T, Buechler P, Reyes M. Atlas-based segmentation of brain tumor images using a markov random field-based tumor growth model and non-rigid registration. Conf Proc IEEE Eng Med Biol Soc. 2010; 2010: 4080-4083. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/21096622
- Chang HH, Valentino DJ. An electrostatic deformable model for medical image segmentation. Computerized Medical Imaging and Graphics. 2008; 32: 22-35. http://bit.ly/2RxuR7N
- Saddique M, Kazmi JH, Qureshi K. A hybrid approach of using symmetry technique for brain tumor segmentation. Computational and Mathematical Methods in Medicine. 2014; 10. http://bit.ly/3aoRaoV
- Abbasi S, Pour FT. A hybrid approach for detection of brain tumor in MRI images. http://bit.ly/2G4FnhD
- Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi. A survey on deep learning in medical image analysis. Med Image Anal. 2017; 42; 60-88. http://bit.ly/2THhODw
- Abdel Maksoud E, Elmogy M, Al Awadi R. Brain tumor segmentation based on a hybrid clustering technique. Egyptian Informatics Journal. 2015; 16: 71-81. http://bit.ly/36anUiz
- Jun Cheng, Wei Huang, Shuangliang Cao, Ru Yang, Wei Yang, Zhaoqiang Yun, et al. Enhanced performance of brain tumor classification via tumor region augmentation and partition. PLoS One. 2015; 10: 140381. http://bit.ly/37dl51w
- Cheng J, Yang W, Huang M, Huang W, Jiang J, Zhou Y, et al. Retrieval of brain tumors by adaptive spatial pooling and fisher vector representation. PLoS One. 2016; 11: 157112. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27273091
- Alfonse M, Abdel Badeeh M. Salem. An automatic classification of brain tumors through mri using support vector machine. Egyptian Computer Science Journal. 2016; 40: 11-21.
- Chunming Li, John C Gore, Christos Davatzikos. Multiplicative intrinsic component optimization (MICO) for MRI bias field estimation and tissue segmentation. Magnetic Resonance Imaging. 2014; 32: 913-923. http://bit.ly/2RL2PWP
- Zacharaki EI, Kanas VG, Davatzikos C. Investigating machine learning techniques for MRI-based classification of brain neoplasms. Int J Comput Assist Radiol Surg. 2011; 6: 821-828. Pubmed: https://www.ncbi.nlm.nih.gov/pubmed/21516321
- Kalavathi P, Prasath VB. Methods on skull stripping of MRI head scan images-a review. Journal of Digital Imaging. 2016; 29: 365-379. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26628083/
- Roslan R, Jamil N, Mahmud R. Skull stripping magnetic resonance images brain images: region growing versus mathematical morphology. International Journal of Computer Information Systems and Industrial Management Applications. 2011; 3: 150-158. http://bit.ly/2sGmxud
- Likar B, Viergever MA, Pernus F. Retrospective correction of MR intensity inhomogeneity by information minimization. IEEE Trans Med Imaging. 2011; 20: 1398-1410. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/11811839
- Jaber Juntu, Jan Sijbers, Dirk Van Dyck, Jan Gielen. Bias field correction for MRI images. 2005; 30: 543-551. http://bit.ly/2G5o9AT
- Dinggang S, Davatzikos C. HAMMER: hierarchical attribute matching mechanism for elastic registration. IEEE Transactions on Medical Imaging. 2002; 21: 1421-1439. http://bit.ly/3aqncRt
- Wang L, Chen Y, Pan X, Hong X, Xia D. Level set segmentation of brain magnetic resonance images based on local Gaussian distribution fitting energy. Journal of Neuroscience Methods. 2010; 188: 316-325. http://bit.ly/2G8rbEe
- Sharma J, Sanfilipo MP, Benedict RH, Weinstock-Guttman B, Munschauer FE 3rd, Bakshi R. Whole-brain atrophy in multiple sclerosis measured by automated versus semiautomated MR imaging segmentation. AJNR Am J Neuroradiol. 2004; 25: 985-996. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15205136
- Rusinek H, de Leon MJ, George AE, Stylopoulos LA, Chandra R, Smith G, et al. Alzheimer disease: measuring loss of cerebral gray matter with MR imaging. Radiology. 1991; 178: 109-114. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/1984287
- Fennema Notestine C, Ozyurt IB, Clark CP, Morris S, Bischoff Grethe A, Bondi MW, et al. Quantitative evaluation of automated skull-stripping methods applied to contemporary and legacy images: effects of diagnosis, bias correction, and slice location. Human brain mapping. 2006; 27: 99-113. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15986433/
- Park JG, Lee C. Skull stripping based on region growing for magnetic resonance brain images. Neuroimage. 2009; 47: 1394-1407. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19389477
- Swiebocka Wiek, J. Skull stripping for MRI images using morphological operators. IFIP International Conference on Computer Information Systems and Industrial Management. 2016; 172-182. http://bit.ly/38rQ0Hh
- Otsu N. A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics. 1979; 9: 62-66. http://bit.ly/2G70IqU
- Yazdani S, Yusof R, Karimian A, Mitsukira Y, Hematian A. Automatic region-based brain classification of MRI-t1 data. PLoS One. 2016; 11: 151326. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27096925/
- Shan ZY, Yue GH, Liu JZ. Automated histogram-based brain segmentation in t1-weighted three-dimensional magnetic resonance head images. Neuroimage. 2002; 17: 1587-1598. PubMed: https://www.ncbi.nlm.nih.gov/pubmed/12414297
- Megha P Arakeri, G Ram Mohana Reddy. Efficient fuzzy clustering based approach to brain tumor segmentation on MR images. http://bit.ly/2G3af29
- Joseph R. brain tumor MRI image segmentation and detection in image processing. International Journal of Research in Engineering and Technology. 2014; 03: 1-5.
- Jin L, Min L, Wang J, Fangxiang W, Liu T. A survey of MRI-based brain tumor segmentation methods. Tsinghua Science and Technology. 2014; 19: 578-595. http://bit.ly/38ohDRC
- Isin A, Direkoglu C, Sah M. Review of MRI-based brain tumor image segmentation using deep learning methods. Procedia Computer Science. 2016; 102: 317-324. http://bit.ly/3asJQZB
- Oliveira G, Varoto R, Cliquet A. Brain tumor segmentation in magnetic resonance images using genetic algorithm clustering and adaboost classifier. http://bit.ly/366Ay21
- Dhanalakshmi P, Kanimozhi T. Automatic segmentation of brain tumor using k-means clustering and its area calculation. International Journal of Advanced Electrical and Electronics Engineering. 2013. http://bit.ly/36cAxJG
- Paul T U, Bandhyopadhyay SK. Segmentation of brain tumor from brain MRI images reintroducing k -means with advanced dual localization method. International Journal of Engineering Research and Applications. 2012; 2: 226-231. http://bit.ly/2TI0pL0
- Dhanachandra N, Manglem K, Chanu YJ. Image segmentation using k -means clustering algorithm and subtractive clustering algorithm. Procedia Computer Science. 2015; 54: 764-771.
- Moon N, Bullitt E, van Leemput K, Gerig G. Automatic brain and tumor segmentation.
- U Baid, S Talbar, S Talbar. Comparative study of k-means, Gaussian mixture model, fuzzy c-means algorithms for brain tumor segmentation. http://bit.ly/2v8Jpn7
- Bezdek J C, Ehrlich R, Full W. FCM: The fuzzy c-means clustering algorithm. Computers & Geosciences. 1984; 10: 191-203. http://bit.ly/30AXDsx
- Pal NR, Bezdek JC. On cluster validity for the fuzzy c-means model. IEEE Transactions on Fuzzy Systems. 1995; 3: 370-379. http://bit.ly/36cElKW
- Zhang YJ. A survey on evaluation methods for image segmentation. Pattern Recognition. 1996; 29: 1335-1346. http://bit.ly/367rPML
- Pont Tuset J, Marques F. Supervised evaluation of image segmentation and object proposal techniques. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2016; 38: 1465-1478. http://bit.ly/2NJLtYX
- Mansoory MS, AA Behboudi M, Refahi S. Performance evaluation of MRI tumor segmentation using clustering algorithms. Arch Med. 2018; 10. http://bit.ly/2ueexRu
- V Malathy, SM Kamali. Brain tumor segmentation from brain magnetic resonance images using clustering algorithm. International Journal of Innovative Technology and Exploring Engineering (IJITEE). 2019; 8: 625-629. http://bit.ly/38p4msc
- Ross MR, Schomer DF, Chappell P, Enzmann DR. MR imaging of head and neck tumors: comparison of T1-weighted contrast-enhanced fat-suppressed images with conventional T2-weighted and fast spin-echo T2-weighted images. AJR Am J Roentgenol. 1994; 163: 173-178. http://bit.ly/2v7PfVJ
- Srinivas B, Rao G S. Unsupervised learning algorithms for MRI brain tumor segmentation. http://bit.ly/2TGg11q
- Kaya IE, Pehlivanli A C, Sekizkardes EG, Ibrikci T. PCA based clustering for brain tumor segmentation of T1w MRI images. Comput Methods Programs Biomed. 2017; 140: 19-28. http://bit.ly/36cGfeA
- Menon N, Ramakrishnan R. Brain tumor segmentation in MRI images using unsupervised artificial bee colony algorithm and FCM clustering. http://bit.ly/38rCR1d
- Militello C, Rundo L, Vitabile S, Russo G, Pisciotta. Gamma Knife treatment planning: MR brain tumor segmentation and volume measurement based on unsupervised Fuzzy C-Means clustering. International Journal of Imaging Systems and Technology. 2015; 25: 213-225.
- Jose A, S Ravi, M Sambath. Brain tumor segmentation using k-means clustering and fuzzy c-means algorithms and its area calculation. International Journal of Innovative Research in Computer and Communication Engineering. 2014; 2: 3496-3501.
- Yadav S, Meshram S. Brain tumor detection using clustering method. International Journal of Computational Engineering Research. 2013; 3: 11-14.
- Vijay J, Subhashini J. An efficient brain tumor detection methodology using K-means clustering algoriftnn. http://bit.ly/38rE3BJ
- Jaware TH, Khanchandani KB. Performance analysis of unsupervised clustering methods for brain tumor segmentation. BRAIN. Broad Research in Artificial Intelligence and Neuroscience. 2013; 4: 1-4.
- Selvakumar J, Lakshmi A, Arivoli T. Brain tumor segmentation and its area calculation in brain MR images using K-mean clustering and Fuzzy C-mean algorithm. http://bit.ly/2RLy9Vp
- Shasidhar M, Raja VS, Kumar BV. MRI brain image segmentation using modified fuzzy c-means clustering algorithm. http://bit.ly/2TG7o7l
- Logeswari T, Karnan M. An improved implementation of brain tumor detection using segmentation based on hierarchical self-organizing map. International Journal of Computer Theory and Engineering. 2010; 591-595.
- Kharrat A, Benamrane N, Messaoud MB, Abid M. Detection of brain tumor in medical images. http://bit.ly/367yu9J
- Ahmed MM, Mohamad DB. Segmentation of brain MR images for tumor extraction by combining k means clustering and perona-malik anisotropic diffusion model. International Journal of Image Processing. 2008; 2: 27-34.
- Selvathi D, Arulmurgan A, Seivi ST, Alagappan S. MRI image segmentation using unsupervised clustering techniques. http://bit.ly/36a92R1
- Capelle AS, Alata O, Fernandez C, Lefevre S, Ferrie JC. Unsupervised segmentation for automatic detection of brain tumors in MRI. http://bit.ly/369oUTS
- Popuri K, Cobzas D, Murtha A, Jagersand M. 3D variational brain tumor segmentation using Dirichlet priors on a clustered feature set. Int J Comput Assist Radiol Surg. 2012; 7: 493-506. http://bit.ly/30BBF8J









Sign up for Article Alerts