Brain Tumor Detection Using Image Processing Techniques
121
different brain segmentation techniques was examined with and without the
skull stripping step. The finding of the study was that segmentation techniques
are significantly more effective when the skull was removed from a brain MRI
image.
In the literature there are publicly available and widely used algorithms for
brain extraction like Brain Extraction Tool (BET) [21], Brain Surface Extrac-
tor (BSE) [22], 3dSkullStrip [23], BridgeBurner (BB) [24], ROBEX (Robust
Brain Extraction) [25], and Graph Cuts (GCUT) [26]. Brief descriptions of
these methods are presented below.
The BET algorithm [21], uses a deformable model that evolves to adapt to
the brain surface. It estimates the intensity threshold between the brain and
non-brain regions. Based on the center of gravity of the head, BET first de-
fines an initial sphere. Then, it enlarges the sphere until it achieves the brain
border. In the BSE algorithm [22], the process starts with anisotropic diffusion
filtering and then edge detection is performed with the 3D Marr Hildret op-
erator. A series of morphological operations are applied to adjust the surface
relative to the brain. 3dSkullStrip [23] is a component of the Analysis of Func-
tional NeuroImages (AFNI) package [27]. It modifies the BET by preventing
leakage into the skull and the inclusion of the ventricles and eyes. The BB [24]
employs a strategy known as thresholding with morphology. BB first locates
a small cubic region in the white matter of the brain, and then it calculates
a window using its mean intensity, which can be utilized to produce a coarse
brain segmentation. A boundary set is created by combining the output of an
edge detector with the surface of the preliminary mask. Then, all connections
between the brain and non-brain tissue are eliminated through morphological
procedures. The ROBEX method [25] accomplishes the final results by com-
bining the discriminative and generative models. The discriminative model is
a random forest classifier, employed to detect the brain boundary and the gen-
erative model is a point distribution model that guarantees that the outcome
is conceivable. The GCUT method [26] compromises of intensity thresholding,
graph cuts and post-processing stages. The first step is to find an appropri-
ate threshold value which falls between the gray matter and cerebro-spinal
fluid intensities and use it to obtain a preliminary mask. Mask will group the
brain and skull, as well as some small interconnections between them. Graph
cuts which is a type of graph theoretic image segmentation technique is uti-
lized to eliminate the narrow connections. In post-processing stage, results are
improved with morphological closing operation.
Besides the above-mentioned algorithms, semi-automatic or fully auto-
matic brain extraction techniques have been introduced in recent years in
order to overcome the drawbacks of manual segmentation. Ilhan and Ilhan [28]
used an opening operation with the structure element disk and a series of pixel
subtraction operators to remove the skull from the brain image. Unintention-
ally, some of the pixels from the tumor with the skull were also removed after
the opening operation. In order to remove the skull from the brain image more
effectively, pixel subtraction was utilized. The image of the brain without the