Article Preview
Top1. Introduction
Medical image segmentation is a method for partitioning or distinguishing organ boundaries in medical images. Although traditionally, radiologists use experience and judgment to perform diagnostics, automatic and semi-automatic image segmentation methods are needed to assist them.
For medical images like CT scans, exposure of radiation is in the range of 1 mSv(millisievert) up to 15 mSv. Within this range, the harm to the patient’s body is minimal. However, prolonged exposure to CT scans can be hazardous. So, MRI images are more preferred for soft tissues (Images, n.d.). As our work is focused on Glioma segmentation, we have preferred MRI data. The dataset used for our experimentation is from the BRATS competition (Menze et al., 2014).
Even though MRI provides details of tissues, they still suffer from weak edges for gliomas. Image segmentation of these weak edges is challenging as thresholding methods may fail (Global thresholding, n.d.). Researchers have used Active methods for segmentation for these kinds of weak boundary problems.
Active methods are based on the philosophy of curve evolution and they are called active, because, the curves in these methods dynamically alter their shape and position while seeking a minimal energy state (Kass et al., 1988). Active methods are categorized into two parts: template-based methods and seed-based methods.
In active methods, there are template-based methods which is used to segment organs with a well-defined structured shape, size, and texture of organs. The template-based methods such as the active shape model (ASM) (Cootes et al., 1995) and the active appearance model (AAM) (Cootes et al., 1998), are based on statistical models of shapes and appearance (texture). Using training data, these models learn about the variation of shape and appearance. The curves of these models alter their shape to fit an example of the object for a new image. So, template-based methods are mainly used in areas like knee segmentation (Vincent et al., 2010), liver segmentation (Heimann et al., 2006).
The seed-based methods need seed point to create contours or initial entry points. The first seed-based method is Kass’s Active contour method (ACM). ACM is based on the energy minimization principle. The total energy is defined by, ‘internal energy’ which is based on the initialization of contour and ‘external energy’ is defined utilizing image properties of the object of interest. Generally, external energy is based on an edge map, which stops on the boundary of the object. The edge map is generated using a constant gradient kernel is used, like Prewitt, Sobel, and Canny. The Gradient kernels are constant kernels and due to which the modelling of weak edges is not satisfactory.
To get a robust segmentation, ACM requires an edge map with strong edges, else the contour boundaries do not stop at ROI during evolution. Though most of the research in literature is towards proposing new methodologies of curve evolution or Internal energy. Research on external energy definition is mostly based on, use of constant kernel edge detector. The constant kernel edge detector fails to capture weak edges, hence there is a need to develop methods to extract a robust edge map.
In this paper, we have proposed a novel biologically vision inspired method based on Adaptive kernels extracted using Independent Component Analysis (ICA) for extracting an edge map. The ICA-based Adaptive kernels in phase congruency methods are used along with Phase congruency, to create robust edge maps. These edge maps when compared with other traditional edge detection techniques to define external energy for GACM, has improved the segmentation performance of Active methods.
For experimentation, we have used the BRATS dataset (Menze et al., 2014) of brain tumors for Glioma (Brain tumor) segmentation. However, as this approach is not based on the Deep learning approach, we have not compared it with current Deep learning methods present on BRATS’ leaderboard.
The remainder of this paper is as follows: The contribution of our paper is explained in1.1. The proposed methodology is discussed 3 and based on the methodology experiments are shown in3.1. Finally, we discussed the conclusions and future scope in section 4.