1. Introduction
With billions of neuronal cells, the human brain presents one of the intricate patterns of structural and neural connectivity in the human organism. The characterization of different sets of the brain, for instance, led to a new multidisciplinary approach in the study of networks [
1,
2,
3]. The brain connectivity characterizes networks of brain regions connected by anatomical traits [
4,
5]. Understanding biological neuronal networks, particularly in the human brain, requires proper knowledge of the network architecture of the whole brain [
6]. Thus, over the past decades, the brain-mapping methods and neuroimaging techniques for the pattern of neuronal networks gained great interest [
7]. In this context, the wide range of quantitative analysis of imaging datasets in the study of the human brain plays an essential role in detecting brain disorders and early diseases [
8].
Magnetic Resonance Imaging (MRI) is a medical imaging modality that attracts attention in biomedical engineering and is known as a safe, non-invasive, non-persistent, and pain-free diagnostic technique. Medical images can be obtained from X-ray radiography, Computed tomography (CT), and other modalities. However, Magnetic Resonance Imaging (MRI) does not cause radiation and employs a uniform magnetic field and Radiofrequency (RF) to expose the human body to gain images of the internal body system. MRI images can be presented with high-quality images in terms of resolution and contrast in 3D and 2D formats. These digital formats give a vast amount of information about internal diseases for soft tissue differentiation and further analysis and classification. MRI can provide detailed information about abnormalities in the soft tissue [
9] that may not be determined by CT or X-ray radiography.
In this paper, we proposed a new method based on convolutional neural network (CNN) and discreate wavelet transform for brain MRI classification. Recent advances in neuroimaging techniques have resulted from complex neurological disorders, often in terms of several challenges in early diagnosis and treatment. On the one hand, these developments have taken place due to continuously produced medical data thanks to tangible progress in automated CAD systems in medical imaging informatics. MRI-based medical images, for instance, are more than pictures; they are data [
10,
11]. On the other hand, most biomedical images show differences in brightness, shape, and texture [
12]. Due to its intrinsic nature, the segmentation process of any medical image is a time-consuming and challenging task [
13]. Therefore, as the images of the human brain fall together with ‘big data’ [
14], there is an increasing demand for an automated image processing to analyze and classify in terms of the latest applications in machine-learning techniques [
15].
Deep learning (DL) is a subfield of machine learning that extends traditional neural network (NN) to models that mainly focus on feature learning. Compared to other DL models, a CNN with a set of algorithms and techniques has become a successful tool in MRI image classification [
16]. An important aspect of CNN in DL is that the necessary features can be learned through directly providing images known as end-to-end strategy, i.e., there is no need to extract information from images first to feed CNN [
17]. CNN has three fundamental mechanisms: a local receptive field, weight sharing, and subsampling, and consists of several layers, including convolutional and pooling layers, and each feature map in a pooling layer is connected to a feature map in a convolutional layer. CNN has been widely used in medical imaging for breast tissue classification and lung nodule detections. Later CNN became very popular in the MR image classification for tumor-like lesions and tissue segmentation and detection and deep cortical and subcortical white matter structures and tissue segmentation [
18].
Quantitative analysis of MRI-based images, in general, plays a vital role in clinical diagnosis for the treatment of neurological diseases. With a high resolution, MRI easily detects signals emitted from normal and abnormal tissue [
19], providing valuable information in distinguishing healthy and diseased brains. Several studies previously have examined in developing machine learning algorithms for MRI-based image segmentation of normal (e.g., white and gray matter) and abnormal brain tissues (e.g., brain tumor) [
13,
15,
20,
21]. Nevertheless, the classification of brain MRI slices as normal and abnormal is still a challenging task [
21]. Developing a robust segmentation method, for instance, is a crucial element in the successful classification of brain MRI images [
13]. This paper deals with the novel classification of MRI data of normal and pathological brain tissues using a robust segmentation method that employs deep learning technique based CNNs. In recent years, CNNs have gained significant interest in medical imaging [
22,
23] and have become more prevalent in image classification methodology.
In image analysis methods, feature extraction is a method of dimension reduction. At some point, the following process concentrates on the extraction of specific features from brain MRI images [
24]. Several methods reported different techniques for feature extraction in image classification, wavelet transform-based techniques [
25] such as Discrete Wavelet Transform (DWT) [
26] and Continues Wavelet Transform (CWT) [
27]. For feature reduction studies, the most used techniques are already available, e.g., Linear Discriminate Analysis (LDA) and Genetic Algorithm (GA) [
28], Independent Component Analysis (ICA) and Principal Component Analysis (PCA) [
29]. Wavelets transform, for instance, has become a prevalent choice for multiple imaging techniques and MRI classification features, thanks to its effective non-stationary signal analysis method [
30]. In this context, we have proposed a novel approach for image classification by integrating wavelet transform to extract features from MRIs.
This work has proposed a novel method based on Discrete Wavelet Transform (DWT) and Convolution Neural Network (CNN) for brain MRI classification. The main contribution is the new assembling of a discrete wavelet transform with the convolutional neural network. The discrete wavelet transform has been used to remove unnecessary detail and make the image more informative and efficient for machine learning algorithms to classify. The reason behind using discrete wavelet transform with a convolutional neural network is that the approximate images returned by discrete wavelet transform have denser information and proficiency for classification than original images.
Most research focused on the classification of brain images as normal or abnormal for abnormal brain MRI studies. In the presence of any pathological appearance, the next stage will be location identification and the medical recommendation. The division of brain MR images into normal and abnormal can be carried out in two ways: (1) using the conventional machine learning models, e.g., artificial neural network, logistic regression, k-nearest neighbors, decision tree, support vector machine, and random forest; (2) using deep learning models, e.g., CNN, stacked autoencoder (SAE), Boltzmann machine (BM), long short-term memory (LSTM), etc. The conventional classification models and the deep learning models have their pros and cons when applied for image classification.
When the normal or standard classifiers are used for the purpose, the major contribution is the feature extraction stage, in which a minimal representation of an image is fed as input to the classifier. In this architecture, some well-known features of images are extracted, reduced, and then given as the input to the classifier. The major drawback associated with this phenomenon is the loss of information during the feature extraction and feature reduction stage. On the other hand, if the feature is not extracted from the image or not even reduced, the classifier is not too powerful to perform the processing of the whole image or a higher number of features efficiently. Hence, a trade-off is required for gaining sufficient information from the images. Eventually, the number of features extracted must neither be too high nor too low for maintaining a fruitful outcome.
Similarly, if deep network models are used to classify brain MRI images, the whole image is given as input to the model for performing the classification task. To process the whole image, the deep network models developed are highly complicated. The complex models add extra processing time and effort to the model processing. In all previous works where deep models have been used for classification, the authors have used the whole image as input to the model, resulting in more processing time as outlined. This drawback of deep models can be overcome if, instead of the whole image, another representation of an image with a smaller size is given as input to the deep model.
Our contribution is three-fold: Firstly, the identification of representation of image adequately enough to represent the whole image without any information loss. After extensive experimentation, the final and summarized representation was the Harr wavelet which is more effective and the simplest wavelet in the wavelet’s family. Other wavelets were also included in the experiments, but their information possessing capability cannot maintain the information. Secondly, compared to previously proposed CNN or other deep models, our approach has a simple CNN architecture due to the reasons mentioned above. Eventually, our proposed model provides simplicity as compared to other deep models with significantly added performance.
The rest of the paper is organized as follows:
Section 2 presents a brief review of the related work.
Section 3 briefly describes the method implemented in this paper.
Section 4 is carried out with the implementation, experimental results, and discussion. Lastly, the conclusion part is presented in
Section 5. The abbreviations with their corresponding descriptions are listed in
Table 1.
2. Related Work
Over the past decades, several studies reported on computer-based neuroimaging techniques for characterization and processing of MRI brain images that have become the tool of choice for the diagnosis of brain disorders and early treatment [
10,
24,
31,
32,
33,
34,
35,
36]. At the same time, however, automated segmentation and classification of normal and pathological brain structures are one of the most challenging tasks [
15]. Nonetheless, numerous approaches have been developed applying machine learning techniques to detect the structural, functional alterations in the human brain; some are described in this section. For instance, the classification of MRI data in image processing is often a costly, laborious, and time-consuming task [
37].
Numerous works have been done toward feature extraction, segmentation, and classification of MRI images to develop different versions of algorithms and deep learning models. Many authors have used conventional techniques integrated with modified algorithms for the preprocessing of MRI images, then following the steps of computer-aided diagnosis (CAD) frameworks to urge the ultimate outputs. All endeavors were aimed at the best models to extend the performance of brain image classification. Many authors applied DWT feature extraction tools to feed a neural network model for MRI classification for image feature extraction purposes. For instance, Chaplot et al. 2006 [
25] employed a DWT feature extraction as an input to ANN and support vector machine (SVM) for brain disorder detection, and Maitra et al. [
38] presented two-stage algorithms of orthogonal DWT for feature extraction and SVM for image classification. Kumar et al. 2017 [
39] proposed a slightly different model, where authors used DWT feature extraction, genetic algorithm principal component analysis (PCA), and SVM classification. PCA was implemented to reduce the number of features, and this hybrid method aimed at MRI tumor classification. El-Dahshan et al. 2010 [
40] used DWT for feature selection and forward back-propagation artificial neural network (FP-ANN) and k nearest neighbor (KNN) classifier tools for MRI brain image classification. A method of clustering Fuzzy C-means (FCM) was utilized by Mohsen et al. 2018 [
41] to image segmentation, and they used DWT for feature extraction and deep neural network for MRI brain tumor classification.
High accuracies in brain MRI classification have been achieved by Wahid et al. [
42], who proposed a method based on statistical moments and probabilistic techniques. Statistical moments have been employed for feature extraction, and ANN has been used for feature reduction. Zahid et al. [
43] proposed another methodology for brain MRI classification using DWT, color moments, and ANN. The DWT method has been used for image decomposition and removed low detail from the image to obtain an approximate small-sized image. A Harr wavelet of three levels of decomposition has been applied to the images. The first three statistical moments are then calculated for each channel and total of 9 features are obtained that have then been further fed to ANN for classification. Slightly different combination was proposed by Amin et al. [
44]. They presented an MRI tumor classification that employs DWT-based image fusion with Daubechies kernel, a global thresholding method for segmenting tumor region and CNN model. A 23 layered CNN architecture utilizes convolutional, batch normalization, rectified linear activation unit (ReLU), down sampling through max pooling, fully connected network, and final output layer softmax to classify normal and pathological brain structures.
Masood et al. [
45] has proposed a method based on fuzzy logic and convolutional neural network for brain tumor detection. In the preprocessed step the image enhancement is used for image segmentation, fuzzy logic has been used for edge detection, and convolutional neural network has been used to classify the brain images into meningioma and non-meningioma. The proposed method is compared with some well-known methods and the results indicate that the proposed method performed well as compared to counterpart algorithms. Muzammil et al. [
46] proposed for improved clinical diagnosis using an innovative multimodal image fusion technique. Obdusami et al. [
47] suggested a method for mild cognitive impairment (MCI), late mild cognitive impairment (LMCI), early mild cognitive impairment (EMCI), and Alzheimer’s diseases (AD) prediction using the finetuned ResNet18 network. The results exhibit that the accuracy of this method is high as compared to conventional methods. A similar approach for pathological brain detection has been proposed by Zhang et al. [
48] based on three components namely wavelet packet Tsallis entropy, extreme learning machine, and java algorithm. It was noted that the proposed method outshines the existing methods.
To make the abnormal image classification process more efficient, Jude et al. 2019 [
49] used simple assignment processes rather than the weight adjustment process to reduce computational complexities in conventional CNN architecture. Utilizing the 2D CNN approach of Simonyan and Zisserman [
50], Kamnitsas et al. [
51] presented a more discriminative 3D CNN model and processed multi-scale parallel convolutional pathways for MRI brain tumor segmentation particularly for large data sets. Pereira et al. [
52] presented their own CNN architecture with the same approach, which employs small, cascaded kernel layers rather than single and bigger ones. This model benefits from fewer weights of the network and results in an effective MRI image segmentation. To discriminate small lesions (<1.5 cc), Liu et al. [
53] proposed a modified version of the CNN algorithm, where authors employed one more sub-path to Kamnitsas et al. CNN model [
51] for the MRI brain metastases segmentation process. To increase classification execution, Togacar et al. [
54] proposed the recursive feature elimination (RFE) embedded CNN model enhanced with hypercolumn technique and supported by networks like AlexNet and VGG-16, and SVM classifiers. Inspired by the residual neural networks, Remedios et al. [
14] presented a 3D CNN architecture for MRI contrast classification and named their model as PhiNet designed for specific diseases like Alzheimer’s, sclerosis, and traumatic injuries.
In addition, some authors presented hybrid models to outperform traditional deep learning techniques. For example, Cinar et al. 2020 [
55] presented a hybrid CNN architecture, which is a sophisticated and modified version of the original Resnet50 CNN model [
56] to increase the accuracy rate. The development occurred by removing the last 5 layers of Resnet 50 and adding 10 different layers. Khan et al. [
57] proposed a hybrid model developed by cascading support vector machine with three pathway CNN models. A different hybrid model approach was presented by Kruthika et al. [
58] for MRI Alzheimer segmentation and classification. This hybrid model consists of fast learning capsule networks (3D CapsNet), 3D autoencoder, and 3D CNN. Chang et al. [
59] proposed a combined CNN model and conditional random fields (CRF) to increase MRI brain image segmentation accuracy. This two-pathway CNN model employs max and min pooling layers on each. Finally, a comparative approach was proposed by Talo et al. [
60]; they compared the following well-known trained CNN models: ResNet50, ResNet35, ResNet18, AlexNet, and VGG 16 to classify MRI images into normal and pathological brain structures, i.e., inflammatory, neoplastic, degenerative and cerebrovascular categories using Harvard Medical school MR image datasets. Different methods along with strengths and weakness have given below in
Table 2.
The current studies have some limitations in one way or another way; some methods are good in accuracy but take a lot of time to compute. Some are fast, but the accuracy of these algorithms is poor. Hence, there is extensive to develop model to tackle these issues.