A Novel Deep Learning Approach for Brain Tumors Classification Using MRI Images

Billions of cells in human mind which makes evaluation very difficult. The central and peripheral nervous systems both include tumors that belong to the genetically, biochemically, and clinically diverse category known as pediatric brain tumors. In terms of incidence and mortality among pediatric malignancies, they are collectively the most frequent solid tumors in children and come in second place after leukemias. In USA, Individuals who are below 20 years of age are diagnosed with 2,200 intracranial pediatric brain tumors annually, accounting for 16.6% of all pediatric cancer diagnoses; 78% of these neoplasms are malignant in origin, while the other 21% are benign or have unknown behavior [1]. About 52% of these malignancies comprise astrocytoma, 21% are primitive neuroectodermal tumors (PNET), 15% are other gliomas, while 9% are ependymomas [1]. Another 6-10% of all pediatric cancers are neuroblastomas (NB), It is the most commonly occurring solid extracranial tumor of the nervous system, which develop in sympathetic ganglia and adrenal medulla that new cases account for around 650 in children per year in USA [2].

These extracranial tumors affect infants under the age of two in close to 50% of cases [3]. There is a possibility that these tumors have an origin that conflicts with some conventional wisdom on tumor development, considering the fact that most Pediatric brain tumors (PBTs) in children occur at a young age. The conventional wisdom that most malignancies take at least ten years to form. In fact, current studies are starting to point to an etiology that involves the interaction of early environmental variables with dysregulated developmental processes.

In total, a hundred and fifty exceptional varieties of BT may be found in humans, which can be grouped into malignant and benign tumors. Benign tumors are discovered at earlier stage in the brain. Brain tumors are commonly referred to as malignant tumors due to the fact they are able to unfold outside the brain [4].

A biopsy is commonly finished to test if the tissue is malignant or benign. Some other places withinside the body, a BT biopsy is commonly acquired no earlier than a definitive mind surgical procedure [5]. In order to gain a correct analysis and keep away from surgical procedure and subjectivity, it's miles essential to broaden an energetic diagnostic device for segmentation of tumor and class from MRI [6].

MRI scans manual evaluation is time-consuming for knowledgeable physicians and radiologists, mainly in complicated instances [7]. In the complex instances regularly require radiologists to examine tumor tissue to adjoining regions and beautify snapshots to enhance perceptual nice earlier than classifying tumor types. This scenario is impractical for huge quantities of data, and guide strategies aren't reproducible. Early detection of BT with excessive predictive accuracy is the maximum critical diagnostic step in an affected person's health [8].

In current years, some automatic structures were used to come across BTs use of MRI scans. Hsieh et al. [9] used specific techniques to categories BT into specific types, namely, place of interest (ROI) identification; extract function; and function choice observed through rating. They mixed neighborhood histology with worldwide histogram moments and quantified the glioma impact use of 107 images, seventy-three low-grade images, and 34 high-quality (glioma) images. Sachdeva et al. [10] describe a computer-aided diagnostic (CAD) gadget that extracts color and texture capabilities from segmented ROIs and makes use of a genetic algorithm (GA) to pick the nice capabilities. However, these methods require sophisticated feature engineering to extract informative hand-craft features for BT classification, which is sometimes very time-consuming. On top of that, their classification performances rely greatly on the quality of the designed hand-craft features.

Deep Learning (DL) has recently demonstrated excellent performance on varied tasks, like classification and segmentation of images in computer vision, and has also been utilized in analysis of medical image [11-13]. This work suggests a new approach using deep learning for fully automatic classification of BTs with weighted contrast-enhanced MRI images. The proposed framework includes an enhancement preprocessing to improve input images quality and a classification phase for   images. Classification. The classification model was built based on (LWCNN) that allows for performing automatic classification in an end-to-end manner.

In this paper, the proposed framework includes an image enhancement and classification steps, as shown in Figure 1. The image enhancement step is employed to improve the contrast of the brain tumor region and the irrelevant background. The extraction & classification step is to automatically extract informative features and classify MRI images. In the subsequent subsections, we will introduce the two steps one by one.

1.png

Figure 1. Methodology overview

The first step in our proposed framework is image enhancement using a variety of image-editing techniques, such as an unsharp filter and a histogram equal. The processing procedures are described as follows. In this stage, we used the identical prior images that had been enhanced using a variety of image-editing techniques, such as the Unsharp filter and Histogram equal, to create a new presentation of the dataset [27], as shown in Figure 2, that presents the images more depth. The techniques of image modification are unsharp filter and histogram equal.

By subtracting from the original image an unsharp or smoothed version of it, the unsharp filter is a fundamental sharpening operator that serves to enhance edges [28, 29]. In order to produce sharper edges, the process of unsharp filtering is frequently employed in the printing and photographic industries. Equation 1 is used in unsharp masking to create an image edge f(x, y) from an input f(x, y).

$g (x, y)=f(x, y)-f \operatorname{smooth} (x, y)$          (1)

The effect of the unsharp filter on the image at the location (x, y) will depend on the surrounding pixel values in the image. If there are edges or details in the image near the location (x, y), the filter will enhance them, making them appear sharper and more defined. However, if the image at that location is already sharp or if there are no edges or details nearby, the filter may have little or no effect.

The histogram equalization is the second step that utilized to modify the intensities of brain tumors MRI images to improve contrast. The technique of representing an image with a limited range of intensity values can effectively enhance the overall contrast of multiple images. This technique can significantly improve the visibility of bone structure in MRI images of brain tumors and also enhance the details of photographs that are either under or over-exposed. This modification leads to a more uniform distribution of intensities on the histogram, utilizing the entire range of available intensities. As a result of this technique, regions with lower local contrast can gain a significant increase in contrast. This is accomplished through histogram equalization approach, which effectively disperses the densely populated intensity values that previously reduced visual contrast [28, 30]. In Figure 2, example of the applied image enhancement approaches to improve the image quality.

2.png

Figure 2. The processing applied to datasets

3.png

Figure 3. Architecture of the designed LWCNN

In the second step of our proposed framework, a lightweight CNN (LWCNN) was designed to improve perform the feature extraction and BT classification with low numbers of parameters. LWCNN is an architecture that is designed to be computationally efficient and have a small memory footprint, making it well-suited for deployment on mobile devices and other resource-constrained platforms.

There are many different architectures for lightweight CNNs, but they typically involve a combination of several design techniques, such as a reduction in model size, spatial reduction, depth wise separable convolution, pruning and quantization.

The goal of these design techniques is to accomplish a good trade-off among model accuracy and computational efficiency, making it possible to run the network on resource-limited devices with minimal latency and power consumption.

In Figure 3, the architecture of LWCNN is shown. It includes five convolutional blocks, each block contains a batch normalization layer, max-pooling layer and Leaky ReLU layer. The details of LWCNN are presented in Table 1.

Table 1. Details of the proposed LWCNN model

LWCNN model is a kind of feed-forward neural network. The utilization of convolutions to capture translation invariance can significantly reduce the parameters numbers required, as the filter becomes independent of position. Convolutional, pooling, and fully linked layers make up the CNN model. These layers carry out a variety of tasks, including feature extraction, dimensionality reduction, and classification. In the convolution process of the forward pass, the filter slides over an input shape and computes a map of activation, which determines the point-wise value of every output [30].

The technique of batch normalization is employed to train deep neural networks effectively by normalizing the contributions to each layer for every mini-batch. In deep learning, constructing a neural network with multiple layers can be challenging due to the architecture of the learning algorithm and the underlying random weights in the initial setup. An underlying reason for this problem could be that resetting weights after each mini-batch can cause a shift in the distribution of inputs to the lower layers of the network, making the learning system continuously chase a moving target. The alteration in the input distribution to network layers caused by weight resetting is commonly known as internal covariate shift. The challenge arises from the fact that the model is updated layer-by-layer in reverse order, from the input to the output, while assuming that the weights in the preceding layers remain constant [31, 32].

To improve performance of deep neural networks, we incorporate batch normalization, which involves scaling the output of each layer. Recall that the term "normalization" refers to rescaling data so that the standard deviation and mean are both zero. The fixed input distributions that would eliminate the negative impacts of internal covariate shift might be achieved by brightening the inputs to individually layer. By normalizing the activations of the previous layer, the assumption that the following layer makes regarding the spread and distribution of inputs during weight update will remain relatively constant, if not entirely unchanged. The differences in the normalized inputs between training and inference might result in observable differences in performance for smaller mini-batches that don't contain an appropriate distribution of models from the training dataset. To address this issue, a modification to the existing technique called Batch Renormalization has been developed. It aims to stabilize the estimates across mini-batches of the variable mean and standard deviation.

Additionally, the benefits of batch normalization include: (a) Hyperparameter tuning has less impact on the model due to the implementation of this technique; (b) smaller internal covariant shift; (c) decreasing the dependence of gradients on the parameters scale or underlying values; (d) At this juncture, weight initialization is marginally less crucial due to the application of this technique; and (e) dropout can be eliminated for regularization.

“ReLU” is one of the most common activation functions utilized in neural networks. Researchers often incorporate this technique between layers to introduce nonlinearity and better handle datasets that are more complex and nonlinear in nature. Figure 4 shows ReLU that can be stated as in Eq. (2):

$f(x, y)=\left\{\begin{array}{c}0.01 \ \forall\ \ x<0 \\ x \ \forall\ \ x>0\end{array}\right.$          (2)

4.png

Figure 4. ReLU

Despite its popularity, particularly on DNN, ReLU has certain drawbacks. ReLU is not continuously differentiable, to start. Gradient cannot be calculated at x=0. Although it is not a major issue, it has a little impact on training effectiveness. All values are reset to zero by ReLU. However, since gradient of zero is zero, neurons arriving at high negative values cannot recover from being trapped at 0, which might be advantageous for sparse input. Since the neuron effectively dies, the issue is referred to as the dying ReLU problem.

5.png

Figure 5. Leaky ReLU

As a result, the network can effectively cease learning and perform poorly. Even though input values are fed to the neural network, the sum of input to the traditional ReLU is always negative despite the weights being appropriately initialized to small random values and large weight updates. The problem is not entirely solved, but current advancements in the ReLU, like the leaky LReLU that enables a greater degree of nonlinearity, allowing for the inclusion of small negative values or facilitating the transition from positive to small negative values. The aim of using the Leaky LReLU is to address these issues into a ReLU function by providing a small negative gradient for negative inputs. Figure 5 and Eq. (3) demonstrate the LReLU and its derivative.

$f(x, y)=\left\{\begin{array}{c}0 \forall\ \ x<0 \\ a x \forall\ \ x>0\end{array}\right.$          (3)

where, alpha is a small constant, typically around 0.01. If pixel value at location (x, y) is positive, then the Leaky ReLU function returns the pixel value unchanged. However, if the pixel value is negative, the function returns alpha times the pixel value.

The effect of the Leaky ReLU function on the image at location (x, y) depends on the pixel value at that location. If pixel value is positive, the function has no effect and returns the pixel value unchanged. If the pixel value is negative, the function applies a small negative slope to the pixel value, which can help prevent the gradient from vanishing during backpropagation in a neural network.

Max pooling selects only the highest activation value, while average pooling downplays the activation by combining the non-highest activations. This issue was addressed by Santurkar et al. [33] who suggested a hybrid strategy that included average pooling and maximum pooling. Dropout [34] and Drop Connect [35] are two major influences on this strategy. Eq. (4) may be used to represent mixed pooling.

$s j=\lambda  \max a i+(1-\lambda) \frac{1}{R j} \sum_{i \in R j} a i$          (4)

The selection between max pooling and average pooling is determined by the value of $\lambda$, which is randomly assigned either a value of 0 or 1. If $\lambda$=0, the operation functions as average pooling, and if $\lambda$=1, it functions as max pooling. The value of $\lambda$ is saved for the forward-propagation phase and is utilized during the backpropagation process.

In this study, we presented framework based on deep learning for fully automatic BT classification with weighted enhanced MRI images. The proposed framework includes an image enhancement step for MRI quality improvement and a lightweight convolutional neural network (LWCNN). We evaluated our proposed method using a publicly accessible set of MRI brain images. It shows accurate classification performances in terms of accuracy, precision, sensitivity, specificity, and F-measure. This demonstrate that the proposed approach holds great promise to be applied in clinical applications. In addition, our studies show that an image enhancement process is helpful for training a deep learning model for BT classification since it allows homogenizing of samples distribution from the training set and from the testing set.