Robust FCM clustering algorithm with combined spatial constraint and membership matrix local information for brain MRI segmentation

https://doi.org/10.1016/j.eswa.2019.113159Get rights and content

Highlights

  • Fuzzy C-Means based clustering algorithm is proposed.

  • Membership matrix local information is introduced to improve fuzzy clustering.

  • New spatial constraint is proposed to reduce effect of noise.

  • Membership matrix local information incorporated with spatial constraint.

  • The proposed algorithm is compared with some well-known methods.

Abstract

This paper presents a robust fuzzy clustering algorithm for the segmentation of brain tissues in magnetic resonance imaging (MRI). The proposed method incorporates context-aware spatial constraint and local information of the membership matrix into the fuzzy c-means (FCM) clustering algorithm. Based upon this approach, an FCM clustering algorithm with joint spatial constraint and membership matrix local information (FCMS-MLI) for brain MRI segmentation is presented, which is more robust against noise and other artifacts. The proposed spatial constraint considers both local spatial and gray-level information adaptively, and to the best of the authors’ knowledge for the first time, the membership matrix local information (MLI) of fuzzy clustering is extracted to be utilized besides the spatial constraint. The proposed method solves two significant drawbacks of spatial constraint-based FCM approaches, which are ineffectiveness in preserving image details as well as confronting noise and intensity non-uniformity (INU) simultaneously. These problems are caused due to utilizing spatial constraints solely. The presented context-aware spatial constraint makes the method robust against a high level of noise while preserving image details. Furthermore, employing the MLI technique improves segmentation results in the presence of noise concurrently with INU. In contrast to spatial constraint-based methods, which just use local information in the image domain, the FCMS-MLI technique utilizes information in both image and coefficient domains. Hence, the proposed method benefits from two different sources of information. Finally, several types of images, including synthetic images, simulated and real brain MR images are utilized to make a comparison among the performances of popular FCMS types (i.e. FCM algorithms with spatial constraint), some new methods and the proposed algorithm. Experimental results prove efficiency and robustness of the FCMS-MLI algorithm confronting different levels of noise and INU.

Introduction

Medical image segmentation is the process of partitioning a digital image into distinct regions to be meaningful and suitable for image analysis and understanding. Analysis of the medical images, as a determinant step of clinical diagnosis and studies, directly depends on accuracy of segmentation results. However, accurate segmentation of brain tissues is a challenging task due to the existence of noise and intensity non-uniformity (INU) in magnetic resonance imaging (MRI). Also, inherent features of brain tissues, including their overlap and closeness of tissues to each other in terms of gray-level, make the task more complicated. Furthermore, unsupervised decision-making is one of the essential functions of expert and intelligent systems, especially in the field of medical image processing. To have such a system, carrying out thorough analyses of images is an indispensable issue. Automatic or unsupervised MRI segmentation is the first and most vital step in the image analyses. Thus, improving the segmentation accuracy of medical images will lead to enhancing the performance of expert and intelligent systems. For this reason, variety of researches have been conducted for MR image segmentation (Rehman, Naqvi, Khan, Khan & Bashir, 2019). The approaches of medical image segmentation can be divided into seven categories: region growing approaches, deformable models, thresholding approaches, artificial neural networks, clustering approaches, atlas-guided approaches and Markov random field models (El-Dahshan, Mohsen, Revett & Salem, 2014). Amongst the mentioned methods, clustering-based techniques became a focus of attention. Clustering procedures are categorized as hard clustering scheme and fuzzy clustering (Yang, Zheng & Lin, 2005). Practical issues of MR images such as poor spatial resolution, low contrast, the overlap of intensities, noise and intensity non-uniformity (INU) make the hard clustering scheme an intricate task. In many image segmentation methods, fuzzy clustering scheme has been widely studied (Gong, Liang, Shi, Ma & Ma, 2013; Panigrahi, Verma & Singh, 2019; Soomro, Munir & Choi, 2019). Fuzzy c-mean (FCM) is the best recognized and powerful method among fuzzy clustering algorithms. FCM faces two drawbacks when used in the segmentation of noise-degraded images (Kouhi, Seyedarabi & Aghagolzadeh, 2011). The first drawback is that the FCM does not incorporate the local spatial information, which makes it sensitive to the imaging artifacts and noises. The second limitation is that the clustering is performed based on intensity of pixel that makes it sensitive to INU (Liew, Yan & Law, 2005; Wang & Bu, 2010).

Numerous researches have been conducted to improve the robustness of the conventional FCM algorithm against noise and INU in MRI segmentation. Still, most of them failed to fulfill these objectives properly. This paper presents a robust FCM-based image segmentation scheme to overcome the mentioned problems, which is more accurate than the existing state of the art methods. The algorithm incorporates a context-aware spatial constraint and the local information of the membership matrix into the standard FCM algorithm. The proposed spatial constraint preserves image details while confronting with moderate and high levels of noise. Furthermore, utilizing the membership matrix local information (MLI) technique besides the spatial information helps the algorithm to be more robust against the simultaneous existence of noise and INU. Also, the method is free of any parameter selection, which controls the effect of spatial constraint.

The rest of the paper is organized as follows. In Section 2, the conventional FCM algorithm and its extensions are presented. The proposed algorithm is explained in Section 3. Experimental and comparison results are given in Section 4 and the paper is concluded in Section 5.

Section snippets

Fuzzy C-Means

Clustering techniques fall into two main categories, namely crisp or fuzzy methods based upon the exclusive belonging of a pattern data to either a particular cluster or many clusters with varying degrees. In comparison with crisp method, FCM is capable of retaining more image information. This algorithm is suggested by Bezdek, Ehrlich and Full (1984) as a substitute for conventional hard clustering. For an image with the size of m × n = N and pixel values of xk, k ∈ {1, …, N} FCM portions the

Our proposed method

In this section, first, our technique for extracting and employing local information of the membership matrix is presented. Then our method is proposed by combining this technique and a new spatial constraint.

Experimental results

The proposed method is evaluated by carrying out different experiments utilizing artificially synthesized images, simulated and real brain MRI. A comparison is made among the performances of the proposed algorithm and previously released FCM-based methods which are: standard FCM, FCM_S, FCM_S1, FCM_S2, GIFP-FCM and FLICM. All mentioned methods are based on incorporating local spatial constraints in the FCM algorithm. Therefore, for evaluating the efficiency of the proposed scheme, several

Conclusion

In this paper, a practical approach is proposed for the unsupervised segmentation of brain MRI tissues. The method is developed based upon modifying the cost function of the classical FCM clustering algorithm by a new spatial constraint (FCMS) that considers the impact of the neighboring pixels. Also, a new technique for utilizing local information of the membership matrix is proposed. The presented MLI technique extracts the local information of the membership matrix and integrates it with the

CRediT authorship contribution statement

Abolfazl Kouhi: Conceptualization, Methodology, Validation, Software, Investigation, Formal analysis, Writing - original draft. Hadi Seyedarabi: Methodology, Validation, Formal analysis, Project administration, Supervision, Writing - review & editing. Ali Aghagolzadeh: Validation, Formal analysis, Project administration, Supervision, Writing - review & editing.

Declaration of Competing Interest

None

References (33)

Cited by (0)

View full text