Lesions Detection of Multiple Sclerosis in 3D Brian MR Images by Using Artificial Immune Systems and Support Vector Machines

Abstract

This paper presents a segmentation method to detect multiple sclerosis (MS) lesions in brain MRI based on the artificial immune systems (AIS) and a support vector machines (SVM). In the first step, AIS is used to segment the three main brain tissues white matter, gray matter, and cerebrospinal fluid. Then the features were extracted and SVM is applied to detect the multiple sclerosis lesions based on SMO training algorithm. The experiments conducted on 3D brain MR images produce satisfying results. KeywoRDS 3D Brain MRI, AIS, Detection, Lesions, Multiple Sclerosis, Segmentation, SMO, SVM INTRoDUCTIoN Multiple sclerosis is an autoimmune chronic disease of the central nervous system especially the brain, the optic nerves and the spinal cord. The symptoms are very variable, numbness of a limb, blurred vision, loss of equilibrium...etc (Xavier et al, 2012). Magnetic resonance (MR) imaging can accurately visualize and locate plaques in both the brain and spinal cord. Depending on the sequences used, they appear white (in technical terms, we speak of “hypersignals”) or black (“hyposignals”). In 2019, more than 2.4 million people suffer from multiple sclerosis .The research is focused on finding innovative treatments to relieve people with MS. The goal of this study is to detect abnormalities of gray matter and white matter in MS from 3D RM Image Many methods have been proposed to automatically segment lesions since manual segmentation requires expert knowledge, is time consuming and is subject to intraand interexpert variability (Vera-Olmos et al, 2016). Veronese et al (Veronese et al, 2013) proposed a fuzzy classification algorithm that uses spatial information for MS lesion segmentation. In addition to spatial information, standard deviation dependent filtering is incorporated into the algorithm to provide better noise immunity. Also, fuzzy logic is adjusted to be more selective on vertical elliptical objects instead of circular objects since most plates are in this form. Saba et al (Saba et al, 2018) presented a method of segmentation of MS lesions beginning with contour detection using the canny algorithm, and then a modified blurred mean c algorithm is applied International Journal of Cognitive Informatics and Natural Intelligence Volume 15 • Issue 2 • April-June 2021 98 to increase the accuracy of the diagnosis. Pre-treatment techniques are applied to get the best result were used, such as the brain extraction tool and binarisation Bassem (Bassem, 2012) proposed a technique for segmentation of Sclerosis lesions by using texture textural features and support vector machines. They used two generic configurable components: a central processing module that locates areas of the brain that may form MS lesions, and a postprocessing module that adds or removes these areas for more accurate data. Based on these configurable modules, single-view segmentation and multiple-section view pipelines are provided to address the limitations found in segmentation results. Khotanlou et al (Khotanlou et al, 2011) proposed a SCPFCM algorithm named based on t membership, typicity and spatial information. Firstly, initial segmentation is applied to T1-w and T2-w images to detect MS lesions. then the non-cerebral tissues are removed by using morphological functions and finally for extraction of MS lesions, the result of the image T1 is used as a mask and compared to the image T2. Ayelet et al (Ayelet et al, 2009) presented a multiscale method to detect lesions in multiple sclerosis based on two phases: segmentation and classification. The first one obtains a hierarchical decomposition of a multichannel anisotropic MRI scans and produces a set of features. These features are used in the second phase via a decision tree to detect lesions at all scales. The authors find that the problem of MS lesions segmentation is still widely open especially for supervised automatic approaches. This motivates to propose an automatic approach for MS lesion detection that uses a supervised learning without an explicit expert intervention. The approach is based on AIS for brain tissue segmentation and SVM with SMO for lesions detection. A number of features to define vector types of specific lesions were calculated and these vector types were used as inputs for SVM. This paper is organized as follows. In section 2, the researchers resent their proposed approach. Section 3 shows the obtained experimental results. Section 4 describes the comparison with a previous work and another proposed method. The final section provides the conclusion of this work. THe PRoPoSeD APPRoACH Automatic segmentation of MS lesions is difficult, as indicated in the previous section due to the large variability of multiple sclerosis lesions. Lesions have deformable shapes, their texture and intensity can vary and their location can also vary from one patient to another. The researchers propose to apply a new segmentation workflow based on a voxel analysis. The method consists of three steps (see Figure 1). For each 3D MR image, the AIS are applied for segmentation of the three main brain tissues white matter, gray matter and cerebrospinal fluid. The authors compute a number of features then the SVM is used for MS lesions segmentation only on the white matter since MS lesions are located in. Brain Tissue Segmentation The researchers started by segmenting the image of the brain MR in the three classes mentioned (grey matter, white matter and cerebrospinal fluid) using the AIS algorithm. Segmentation by Artificial Immune Systems (AIS) Artificial immune systems is a model that encompasses both, mathematics and biological principles, as the natural immune system offers interesting features like memory and learning that will be useful for solving problems (Tavana et al, 2016). International Journal of Cognitive Informatics and Natural Intelligence Volume 15 • Issue 2 • April-June 2021 99 Learning This phase is to find the memory cells (voxels) representing the regions in this study area and then make the classification using the algorithm of CLONCLAS that uses the principles of artificial clonal selection. The elements which are used in this algorithm are (Komaki et al, 2016): • Antibody: or samples represents the basis of training that will recognize the image of antigens. • Antigen: represents the basis of examples for which we want to determine the class. • Affinity: affinity in immune systems is the measure of similarity between the antibody and the • Antigen: the latter two are represented by a point. Affinity is the distance between these two points, in this work, the Euclidean distance is used. • Memory cell: represents the best antibodies found for the class. Learning occurs class by class by the CLONAG algorithm (Komaki et al, 2016): 1. Training samples are considered a priori to antibody (Ab). One of these samples randomly drawn is likened to an antigen (Ag). Through the Euclidean distance, they calculated affinity of this Ag with all Abs of the class. 2. Abs voxels are ranked in descending order according to their affinity compared to the Ag considered. The first N voxels will be selected to undergo a cloning while preserving the first voxel to form the memory cell Mc class match. Figure 1. Flowchart of the proposed approach for MS lesions segmentation International Journal of Cognitive Informatics and Natural Intelligence Volume 15 • Issue 2 • April-June 2021 100 3. Clone the n selected voxels i in proportion to their affinity. The clones number of a voxel is even higher than the affinity of the voxel is high. This number is calculated as follows: The number of clones for each member: Nc = round (β* (n/I)) (1) With Nc is the number of clones of an element, β is the cloning coefficient, I is the position of the element to be cloned, and round is a function that rounds a real number to an integer. The total number of clones (Erik et al, 2012):