Combining Local Binary Pattern with Stationary Wavelet Transform for Multiclass Classification of Brain Disorders

The diagnosis of different disorders is the most important step in deciding the treatment for the disease. Recently, brain disorders that strike at different phases of human life are among society’s most severe pathological states. Various imaging techniques and methods have emerged for brain disease identification. The approaches using wavelets are widely used in different image processing and diagnosis applications [1, 2]. The structural MRI modality presents brain disorders non-invasively and with better contrast [3]. However, the capacious MR imaging data can be arduous to interpret manually. Therefore, computer aided detection (CAD) systems are being developed to help physicians in the interpretation of medical images, resulting in improved diagnostic accuracy. Various CAD systems have been proposed by researchers for classifying normal and abnormal MR brain images. Zhang et al. [4] used wavelet coefficients and BPNN classifier for classification of brain disorders, and the study [5] developed an algorithm using ripplet transform (RT) along with multi-scale geometric analysis. They utilized the least square support vector machine for classification. El-Dahshan et al. [6] segmented MR images using feedback pulse-coupled neural networks. The study [7] achieved 97.78% classification accuracy over 90 MR images using wavelet energy as a feature and support vector machine (SVM) classifier. Zhang et al. [8] calculated wavelet entropy of approximation wavelet coefficients and used BBO and PSO to optimize NN for classification. Zhang et al. [9] extracted SWT coefficients from MR images. The features were classified using a variant of the SVM classifier. Wang et al. [10] utilized a twin SVM classifier to classify the extracted dual-tree complex wavelet transform features. Wang et al. [11] reduced the stationary wavelet transform (SWT) coefficients using PCA. They proposed a classifier using variants of feed-forward neural network (FNN) based on hybridization methods of PSO and ABC. Zhang et al. [12] classified the entropies of wavelet packet transform coefficients using generalized eigenvalue proximal SVM (GEPSVM). In the study [13], wavelet entropy (WE) feature was calculated from approximation and detail wavelet coefficients. Features were selected using binary particle swarm optimization (BPSO) and its variants. Nayak et al. [14] employed probabilistic PCA (PPCA) to reduce the 2D DWT coefficients and perform classification using Adaboost algorithm with random forest (ADBRF). Nayak et al. [15] used energy and entropy values, calculated from SWT coefficients, as features. They employed Adaboost algorithm with SVM classifier. Nayak et al. [16] classified MR brain images using curvelet features and three kernels in the least squares SVM classifier. They described promising results. Khalil et al. [17] employed various features, such as local binary pattern (LBP), gray level co-occurrence matrix (GLCM), and histogram of oriented gradient (HOG). They combined the classifiers of these feature extraction techniques using different fusion operators. Gudigar et al. [18] compared three different multi-resolution analysis techniques to classify a dataset of 612 MR brain images using SVM as a classifier and achieved 97.38% accuracy with shearlet coefficients.

Nayak et al. [19] proposed transfer learning approach and Tamilarasiand and Gopinathan [20] proposed Inception Architecture for brain abnormalities detection. Hu et al. [21] explored the approach of fuzzy system for feature recognition and predicting the brain disease. Shanker and Bhattacharya [22] segmented region of interest of MR brain images using their developed Pulse-Coupled Neural Network and classified the brain images using Twin Support Vector Machine. Convolutional neural networks are being widely used in biomedical image examination [23]. Takrouni and Douik [24] combined Curvelet Pooling (CP) and Adam gradient calculation method for improving classification accuracy of brain pathological images. Dora et al. [25] suggested multiple kernel-based convolutional neural network (MK-CNN) to classify the pathological brain images, which is a flexible and high-capacity approach. Zhou et al. [26] addressed the dual challenges of data privacy and MRI brain tumor disease detection using an innovative approach of leveraging Federated Learning (FL). Akter et al. [27] proposed a model using deep Convolutional Neural Network (CNN) for automatic brain image classification and a U-Net-based segmentation. Kale et al. [28] suggested new feature extraction method using local binary pattern and steerable pyramid (SP) and then MR brain images were classified using neural network.

The literature review reveals that wavelet-related transforms are the most widely used feature extraction methods in CAD systems. The multi-resolution property of these transforms retains significant information about images. PCA is the extensively employed feature reduction method whereas SVM, ANN, and KNN are used as classifiers in most of the work. In these works, promising accuracies are achieved. However, the number of images used in these methods is small; therefore, Gudigar et al. [18] introduced a new dataset DS:612 compromising 612 MR images.

In this paper, we compared the classifier performance results using three features individually. The following is the contribution of this work:

• MR image is encoded in local binary pattern using LBP. The wavelet coefficients of LBP coded image are extracted using SWT. The statistical feature of SWT coefficients give multi scale directional representation for brain abnormalities encoded in the LBP transformed image.

• 612 images of dataset DS:612 are rotated and translated by different angles and coordinates respectively to obtain dataset DS:1836 of 1836 images.

• Energy, entropy and standard deviation values of each multi scale sub band of LBP-SWT decomposed image are calculated.

• The performance measures of the back propagation neural network classifier are compared using each feature individually over all datasets.

By performing cross-validation with the Harvard Medical School Datasets [29], the proposed method achieved promising performance for binary and multi-class disease classification. The article is arranged as follow. Section 2 presents material. Methodology of presented system is delineated in Section 3. Section 4 highlights the results along with discussion, leading to conclusion in Section 5.

The methodology of proposed feature extraction and classification system is summarized in Figure 2.

2.png

Figure 2. Proposed feature extraction and classification method

3.1 Local binary pattern (LBP)

LBP operator is applied to each MR image. A small cell size of 3 × 3 is selected, as local abnormalities can be detected effectively with a smaller size.

LBP computes binary code of each pixel. In 3 × 3 neighbourhood, the LBP code of central pixel is calculated by following equations,

$L B P_{P, R}=\sum_{p=0}^{p-1}\left(s\left(g_p-g_c\right)\right) 2^p$                      (1)

$\begin{gathered}s(x)=0, \quad \text { for } x<0 \\ 1, \quad \text { for } x>=0\end{gathered}$                (2)

where, (P, R) specifies a neighbourhood of P, equally spaced points on a circle of radius R, s(x) is the thresholding function, g_c is the centre pixel and p = 0, 1,…., p−1 are the pixels in the neighbourhood.

3.2 Stationary wavelet transforms (SWT)

The lack of translation invariance of the DWT is overcome in SWT. This is accomplished by removing the down samplers and up samplers in the DWT and up-sampling the filter coefficients. Due to this property, it is used in various applications despite being redundant [9, 15]. Figure 3 describes decomposition steps of image using 2D-SWT. At level j, 2D-SWT decomposes approximation coefficients (cA_j) in four components namely the approximation coefficients (cA_j+1), and the details (horizontal (cD^h_j+1), vertical (cD^v_j+1), and diagonal (cD^d_j+1)) in three orientations. The merits of SWT are shift invariance and it also removes the limitation of size constrains of standard DWT (power of 2). As SWT does not use decimation operation, it can be applied to image of any size.

3.png

Figure 3. Decomposition steps of image using SWT

3.3 LBP-SWT approach

Wavelet transform (WT) has been used in combination with LBP to describe texture images [30-33]. In most of the applications, wavelet-decomposed sub-bands are used from which LBP features are extracted. In this paper, LBP [34] is used with SWT to provide a robust brain disease descriptor. In this work, the wavelet used is Symlet2. Its smaller compact support is useful in detecting smaller changes in the image. SWT is applied on LBP-transformed image and decomposition is carried out to level 8. The 4 sub-bands namely approximation, horizontal, vertical, and diagonal of each decomposition level contribute to a total of 32 sub-bands. All 32 sub-bands are considered for feature calculation. The LBP-SWT approach combines the power of LBP as a local texture descriptor and the power of the SWT in detecting spatial changes.

3.4 Feature extraction

Feature represents measure of image structure. In this work, three features namely energy, Shannon entropy, and standard deviation are calculated from 32 multi-scale directional sub-bands of the LBP-SWT image. The features are tabulated in Table 2. In Table 2, X_k (i, j) indicates the kth sub-band and its dimension is M × N. k varies from 1 to 8. µk indicates mean of kth sub-band. Energy quantifies the strength of the image details in different sub-bands. As an individual feature, it has been shown to be effective in analysis of MR brain images [7, 35]. Entropy reflects randomness and in turn complexity within brain image. In the classification of normal and abnormal MR brain images [15, 36, 37], entropy feature has been proved to be competent. Standard deviation provides dispersion measure of coefficient values around mean. It has been used in different applications [38, 39]. The discriminating power of these features varies with the different resolution levels of the image. Such a multi-scale resolution strategy is useful for pattern recognition, and classification [40]. This work, evaluates the competency of these three features in differentiating brain disorders.

Table 2. Feature formula

3.5 Back propagation neural network (BPNN) Classifier

To perform the classification, a supervised learning method, BPNN is used. It is biologically inspired [41]. It can generalize to new using past examples. With these properties, it has achieved considerable success when it is used to solve multivariate, difficult, non-linear, and diverse problems [41-43]. Ojha et al. [44] have given a very good review of the neural network. Optimal selection of hidden layers and also hidden neurons is a crucial part of NN architecture design as it can lead to under-fitting or over-fitting issues. In this work, to overcome the generalization and over-fitting problem of BPNN, the Bayesian regularization method is proposed along with stratified cross-validation. For binary classification, one hidden layer with 5 neurons is used. In multi-class classification, the increase in the number of classes to distinguish, results in increased training equations. So, BPNN with 5 hidden neurons can cause under-fitting issues. To overcome this issue, one hidden layer with 15 neurons is used in multi-class classification for the first three datasets.