Please send me a matlab code for skull stripping using graph cut method. my id is kpaswathikp[at]gmail.com


Regards,
Aswathi

skull stripping using graph cuts matlab code

Abstract
In this paper, we propose a novel technique for skull stripping of infant (neonatal) brain magnetic resonance images using prior shape information within a graph cut framework. Skull stripping plays an important role in brain image analysis and is a major challenge for neonatal brain images. Popular methods like the brain surface extractor (BSE) and brain extraction tool (BET) do not produce satisfactory results for neonatal images due to poor tissue contrast, weak boundaries between brain and non-brain regions, and low spatial resolution. Inclusion of prior shape information helps in accurate identification of brain and non-brain tissues. Prior shape information is obtained from a set of labeled training images. The probability of a pixel belonging to the brain is obtained from the prior shape mask and included in the penalty term of the cost function. An extra smoothness term is based on gradient information that helps identify the weak boundaries between the brain and non-brain region. Experimental results on real neonatal brain images show that compared to BET, BSE, and other methods, our method achieves superior segmentation performance for neonatal brain images and comparable performance for adult brain images.

Introduction
Skull stripping is an important step in brain image analysis and refers to the removal of the scalp, skull, dura, eyes, and other extraneous regions. Tissue classification, registration, volumetric analysis of the brain, and brain surface reconstruction all depend upon accurate skull stripping. Any accidental removal of brain tissues is damaging because it cannot be reversed in later processing stages. Popular skull stripping methods include region-based approaches [1, 2], boundary-based techniques [3, 4], and hybrid methods [5, 6]. In brain image analysis terminology, brain extraction refers to skull stripping, while brain segmentation refers to classification of the brain into different tissues (e.g., white matter, WM; gray matter, GM; and cerebrospinal fluid, CSF). In this paper, we propose a novel method using graph cuts that incorporates prior shape information for skull stripping of infant (neonatal) brain magnetic resonance (MR) images.

Segmentation of neonatal brain MRI is important for the study and treatment of brain injury and disorder due to prematurity. Shortly after an infant is born, neurodevelopment includes critically important maturational processes which may be measured quantitatively by brain imaging. Brain tissue volumes have been shown to change with age [7]. In neonatal brain segmentation, tissue classes apart from WM, GM, and CSF are identified to characterize brain development. In such a scenario, accurate skull stripping assumes increased significance.

Most of the skull stripping methods are designed to work with T1-weighted images as it is the most popular modality for brain MRI due to its superior contrast over other modalities like T2 or FLAIR. Most state-of-the-art skull stripping algorithms have been developed for adult brain MR images. When used on neonatal images, these algorithms do not obtain the high segmentation accuracy of adult brain volumes. Figure 1 shows an example image of adult and neonatal brain MRI to illustrate their differences. Adult brain MRI have a well-defined boundary between the brain and skull, while in neonatal brain MRI, the brain and skull are not easily separable. Unlike neonatal volumes, in adult brain MRI, different tissues are quite clearly defined which provides more information for accurate skull stripping.

Neonatal brain extraction has some unique challenges compared to adult brain volumes. The neonatal data are characterized by poor image quality due to their inherently low spatial resolution, insufficient tissue contrast, and ambiguous tissue intensity distributions [8]. Two very popular skull stripping methods are the brain surface extractor (BSE) [1] and the brain extraction tool (BET) [3]. BSE uses a combination of anisotropic-diffusion filters, Marr-Hildreth edge detectors, and morphological operators to separate brain and non-brain tissues, but needs parameter tuning for specific images. BET initializes a spherical mesh around the center of gravity of the brain and uses a deformable model. Internal and external forces push the initial volume to the brain boundary. BET is fast and relatively insensitive to parameter settings, but can produce areas wrongly identified as the brain. 3dSkullStrip, part of the AFNI package [9], is a modified version of BET. It is adapted to avoid segmentation of eyes and ventricles, reduce leakage into the skull, and use data outside the surface to guide its evolution. The watershed algorithm (WAT) [2] is an intensity-based approach that relies on preflooding, and the basin represents the brain. The hybrid watershed algorithm (HWA) [6] which forms part of FreeSurfer software [10] exhibits greater robustness than other methods, by combining a watershed algorithm, a deformable surface, and a probabilistic atlas. The watershed algorithm makes an initial estimate of the mask by assuming connectivity of the white matter, and a statistical atlas is used to guide the evolution of a smooth surface and refine the mask.

Other approaches to skull stripping segment the brain using intensity thresholds followed by morphological operations to cut narrow connections between brain and non-brain regions [11–13]. Such operations can only remove very narrow connections (weakly connected regions). To overcome this limitation, Sadananthan et al. [14] propose a graph cut-based approach to position cuts for isolating and removing dura. The method in [15] removes narrow connections using distance transforms followed by watershed algorithm (DWAT). In [4], active contours were used to fit the brain where the curve is embedded in a higher dimensional function and locally adapted to reduce sensitivity to bias field. Zeng et al. [16] proposed a system of two level sets whose zero level curves represent the inner and outer boundaries of the cortex. Rehm et al. in [17] use a hierarchy of masks from different models to form a consensus mask for brain segmentation. A learning-based brain extraction system (ROBEX) was introduced in [18] which combines a discriminative and a generative model for brain extraction. The discriminative model is a Random Forest classifier trained to detect the brain boundary while the generative model is a point distribution model to ensure a plausible result. For a new image, the generative model is used to find the contour with the highest likelihood according to the discriminative model. The contour is then refined using graph cuts to obtain the final segmentation. In [19], a method is proposed for segmentation of pediatric brain tumors. It combines probabilistic boosting trees (PBT) and lower-level segmentation via graph cuts. Tu et al. in [20] propose a method using PBTs for automated extraction of major sulci from brain MRI.

In this paper, we propose a method for separating the brain and non-brain from neonatal brain MR images. Graph cuts are used to segment the 3D volume and also incorporate prior shape information. The method in [14] first extracts a rough estimate of the brain and uses graph cuts to refine the segmentation. However, we use graph cuts to extract the whole brain. Prior shape information is included from labeled training data. This paper makes the following contributions: (1) graph cuts are used exclusively for brain extraction or skull stripping. (2) A simple approach is proposed to include shape information with graph cuts by constructing a prior shape atlas from manually labeled segmentations. (3) Gradient information from labeled training data is used to formulate the smoothness term which increases segmentation accuracy. A description of our method is given in “Materials and Methods.” Comparative results of our algorithms with three methods, i.e., BET, BSE, and graph cut (GC) [14], are presented in “Experiments and Results.” Finally, we present a brief discussion (“Discussion” section) and conclusion (“Conclusion” section).

Neonatal Brain Extraction

There are many works related to neonatal brain segmentation [21–23] which make use of conventional brain extraction tools (BSE or BET). However, neonatal brain extraction has some unique challenges not observed in adult brain MRI. As infants grow old, brain structures develop leading to less complications in brain extraction. For example, neonatal brain MRI have low tissue contrast and low image resolution. Consequently, even popular and freely available software like BET and BSE fail to accurately extract the brain. In spite of manual tuning of parameters, parts of the skull are still left connected to the brain due to the weak boundary between the brain and the skull. Neonatal MRI have very low contrast-to-noise ratio posing difficulty in segmenting regions exhibiting partial volume effect. Brain segmentation methods like [8] are able to use BSE and BET for brain extraction because their datasets consist of 1-year- and 2-year-old brain images. In older infants, brain structure is sufficiently formed for BET and BSE to give accurate segmentation results. In our datasets, the maximum age of the infants was 1 month where brain structures are not properly formed, and low-level information alone does not provide accurate extraction results.

Graph Cuts and Importance of Prior Shape Information

We choose to use graph cuts optimization because of its following advantages:

Graph cuts can easily find the global optima of Markov random field (MRF)-based energy function with two labels [24]. MRF-based energy functions are suitable for problems where the solution is represented as a set of labels (e.g., segmentation). MRFs enable the inclusion of context-dependent information from the pixel neighborhood and allow for a regularized solution. This ensures that neighboring pixels take similar segmentation labels and avoid isolated patches of incorrectly labeled pixels. Since our segmentation had two labels (brain and non-brain), graph cuts give a globally optimal solution in quick time.
Graph cuts are not sensitive to the initialization of labels. It gives a globally optimal result irrespective of the initialization and does not get trapped in local minima [24]. This provides a distinct advantage over level sets.
Many knowledge-based algorithms have been developed for neonatal brain image segmentation [8, 22, 25–28] under the guidance of an atlas encoding prior knowledge of anatomical structures, their spatial locations, shapes, and their spatial relationships. Prastawa et al. [22] generate an atlas by averaging three semi-automatic segmented neonatal brain images registered with affine transformation. Song et al. [28] built an unbiased atlas from nine out of ten neonates in a leave-one-out manner with diffeomorphic flow registration. Xue et al. in [29] use multiple age-specific atlases in an expectation-maximization framework for tissue segmentation. In MRF energy functions, contextual information is incorporated from the immediate neighborhood of a pixel. As a result, incorporating shape information in graphs is a challenging task because reliable shape information is obtained from a set of points over a larger neighborhood. The penalty term of the MRF energy is calculated for every pixel while the smoothness term considers inter-pixel interactions in the immediate neighborhood. Thus, the effectiveness of smoothness cost in including prior shape information is limited. The data penalty can be used for including prior shape information but needs a lot of training data to construct a generalized prior shape model.

The first works to use prior shape information in graph cuts were [30, 31]. In [30], the zero level set function of a shape template was used with the smoothness term to favor a segmentation close to the prior shape. Slabaugh et al. in [31] used an elliptical shape prior, under the assumption that many objects can be modeled as ellipses. They apply many iterations as a pre-initialized binary mask is updated to get the final segmentation. Vu et al. [32] use a discrete version of shape distance functions to segment multiple objects, which can be cumbersome. A flux-maximization approach was used in [33], while in [34], the smoothness cost was modified to include star shape priors. Although there are not many works that use prior shape knowledge exclusively for brain extraction, some methods have used shape information to segment parts of the human brain like the corpus callosum [30, 35] and cerebellum [36].