Segmentation of MRI Brain Tumor Image using Optimization based Deep Convolutional Neural networks (DCNN)

1 Introduction

Image analysis is a primary function in medical diagnosis and test [1, 2]. The major health issues in the world are brain irregularities like tumor and brain stroke. It involves the public irrespective of femininity and ecology. Generally, the brain abnormality is diagnosed on a predefined procedure constructed by experienced doctors. The brain images kept under a control environment. Once the recording process is finished, the required information from the patient is executed in a screening procedure by guided or automatic methods to review the causes and condition of severity. The evaluation process is executed with the patient to control and cure the abnormalities after the screening process [3].

Generally, the state of the brain is calculated based on channel and multiple channel EEG signals. It is evidenced by the outside electrodes or the brain images documented by MRI imaging advances [4,5,6,7]. Earlier lessons prove that the in-sequence of the brain is necessary to categorize and confine the abnormalities as contrasted to the strong brain signs [8]. Other methods can be applied to plot brain signals and use images to put reason and area of brain abnormalities. In content, a soft computing directed actions be used to observe the CT and MRI brain images [9, 10]. MRI grasps a variety of modalities style such as T1 T1C, T2 and DW [11, 12, 14]. The observable of irregularity in T2 and Flair are better and aimed for testing. To assess the brain anomaly, some soft computing methods are implemented. Some mechanisms also propose to apply a single step or two step process to segment the irregular part for a selected MRI piece [11, 15, 16]. Many of these processes are fine for few modalities, for example flair and T2. The heuristic algorithms supporting two steps occurrence are widely accepted to challenge this pitfall. It is inspected better in the brain abnormalities verification by MRI of Varity of modalities.

Palini et al. [8] implements otsus thresholding procedurean advance soft-computingsoft computing to dividesion the flexible brain tissues [4, 5]. Chaddad [17] applied Gausian mixture method to extract features of MRI brain tumor. Day et al. [6] used genetic algorithm to eliminate brain MRI noise

A specified estimation of advance brain tumor segmentation is discussed by Rajinikantha and Satapthy [9]. The authors realized a variety of heuristic measures to split the irregularities of brain MRI recorded by different modalities. They state that image a fusion technique is needed to improve the correctness of brain MRI assessment [13]. Amin et al. [4, 5] suggested a unique advance of deep a learning approach to mine the irregular parts of the brain MRI. Every deep learning and machine learning approach has advantages and disadvantages. The assessment process of MRI brain images should well organized and consistent to take out and asses the ROI of brain images [19]. Already a remarkable study has been done to implement evolutionary/swarm to assess brain MRI segmentations [9, 13]. The work done by Author [20] adopted a new tactic using Jaya algorithm. This work confirms that Jaya algorithm is easy compared to previous heuristic algorithms. The study used the chan-vese technique to extract the tumor part in processed brain images. The quality of execution is verified by a better assessment among the mined tumor part and the images in GT [9]. Author [28] test the optimization of CNN using two different datasets i.e., Dermquest and DermIS. The outcomes can be balanced with different techniques like AlexNet, Ordinary CNN, VGG-16, and ResNet etc. Their proposed method achieved better performance.

In this paper [29], a DCNN Model is considered. The Adapt Ahead optimization algorithm enriches the accuracy and decreases losing functions from the training procedure of the model. In this version of training procedure, version-based optimisation algorithms are presented to test the loss in evaluation contrasted to the optimization algorithm that was proposed. This algorithm is more suitable than other marketing algorithms that are based on the BRATS Dataset. From the reviewed literature, we noted that neural system models give promising outcomes in almost each and every field. Convolutional neural networks also perform well in image processing. Optimization algorithms have been employed in these neural network models to set optimized values for model parameters. This combination enhances the performance of the model by converging it faster. As such, we got motivated to use this hybrid composition of optimization algorithms with deep learning neural system for segmentation of brain images. In this context, we propose a hybrid framework to extract the tumor section by optimization based on the DCNN model to mine the irregular part in the (2D) slices of the MRI brain images. We consider swarm-based optimization algorithms like genetic algorithm (GA), particle swarm optimization (PSO), gray wolf optimization (GWO) for comparison [20]. These optimization algorithms have proved their efficacy in almost all research domains.

The organization of rest of the paper is described as follows. Section 2 elaborates the preliminaries and details of optimization algorithms. The working principle of the proposed method i.e., optimization based DCNN model is elaborated in section 3. Section 4 focuses on experiments and discussing the result. Finally, the conclusion of this study is presented in section 5.

2 Material and methods

Neural network presents notable performance in various research works. To handle complex applications, researchers have assumed the development of typical neural networks. The most important and advanced neural network is the deep learning neural network. In this work, we use optimization algorithm in a convolution neural network, like particle swarm optimization (PSO), genetic algorithm (GA), gray wolf optimization (GWO) and whale optimization algorithm (WOA). The use of these optimization techniques in the training process is to optimize the output of the result vector on DCNN and advance segmentation accuracy. This study the MRI brain tumor images from BTARTS 2015 dataset. This dataset contains 220 subjects with high grade and 54 subjects with low grade gliomas. For testing, 53 subjects with mixed high- and low-grade gliomas were used.

3 Proposed Hybrid Frameworks

Although there are variety of research outcomes found on the brain tumor image segmentation and classification process, the outcomes are far way from being fine gained. Individual research techniques are recommended for the new height of excellence. The entire brain tumor recognition schemes rely on suitable pre-processing methods for achieving consistency and quality correctness. A regular segmentation formula for MRI brain tumor using DCNN should facilitate handling the issues that have been discussed in the literature for quality input for brain MRI segmentation stage. The complete automatic model for segmentation approach for MRI brain tumor by DCNN uses the following steps as well as the framework depicted in Figure 1.

1. Image pre-processing: It consist of following three steps

   1. Read image

   2. Resize image

   3. Remove noise

2. Segmentation of image using optimization based DCNN framework and filtering of ROI object based on histogram.

3. Validation has to be done with ground truth and then the final decision has to be made.

Figure 1

A general deep learning framework for automatic brain tumor segmentation

Initially we read the image, resize the image and deploy MICO bias correction method to remove noise from brain image slices. Optimized DNN algorithm is used in the subsequent step to identify brain tumor.

This procedure at first takes a 2D MRI slices as the experiment picture to be processed and the pre-processing phase eliminates noise and prepares a correct image using MICO scheme. This scheme is a proven error minimization technique. Rajinikant and Ratapthy [9] focused on skull removal which is required to design an automatic bug analysis scheme.

3.1 Pre-processing stages with MICO noise removal method in MRI Brain Image

Intensity at homogeneity could be complex in brain MRI brain tumor segmentation. From the tests of prejudice field improvement, we used the MICO method [21] to assess the bias field and cyst segmentation that depends in 2 multiplicative parts on breakdown of an MRI image. This is an energy minimization strategy to streamline joint bias field estimation which is exhibited in Figure 2. BRATS utilize multi-institutional pre-operative MRIs and targets the segmentation of heterogeneous (in appearance, shape, and histology) mind tumors, particularly gliomas. The MRI picture set of each individual includes many different strings including T1, T2-weighted, T1-weighted, along with FLAIR, together using an annotation of both edema, improving tumor, together side tumefaction. For more details about the BRATS statistics set, we employed the BRATS 2015 schooling data set which is freely available, whilst the BRATS difficulty is stored.

Figure 2

MICO bias corrections

3.2 Optimization based CNN model for brain MRI image segmentation

In CNN based image segmentation; a segment of an image has to feed as input to it which in turn levels the pixel. The CNN cannot process the entire image at once. So, the image is scanned and each time several pixels are filtered till the entire image is mapped. In our work, a convolutional neural network with optimization framework is proposed as depicted in Figure 3. In the proposed framework, the output image of the CNN converges to the expected image by adjusting the network parameters of the CNN with the use of optimization algorithms like WOA, GWO, GA, and PSO. We have focused on two popular optimization techniques to optimize parameters of CNN for the segmentation process using the whale optimization algorithm and Wolf optimization algorithm. The objective function's fitness value is calculated by comparing the output image of the CNN and manual doctors prescribed segmented image (GT).

Figure 3

Proposed Framework: Optimization based CNN

The optimization algorithms evaluate their fitness values to obtain the optimized learning parameter of the CNN model.

The working procedure of proposed research work consists of following steps:

Training Phase:

1. Initialization – Initialize the learning coefficient of CNN. Consider the weight vector of CNN as a swarm of optimization algorithm. Initial swarm group size is equal to 50. Number of iterations to be set, in our case we have set it to 50 or till converges.

2. Training process of CNN will start with each swarm optimization algorithm and the first image of the input data set.

3. In this process, once all the swarm or weight vectors have been applied, then it needs to be updated using the updating equation of wolf optimization algorithm.

4. Similarly, all the training images will be applied one by one with procedure mentioned in step 3 till the loss function of CNN model becomes negligible or converges.

5. Once the loss function converges, then we need to freeze the parameters of the CNN model by selecting the individual of the population obtained in the final iteration.

Testing Phase:

1. Once the training procedure with above steps is completed along with 60000 MRI brain image training samples, the testing of the model is carried out with the rest of the 1000 tests brain images.

2. Each of the test brain image sample has to be applied to the trained model with the frozen parameter values obtained from the training phase.

3. Output of the proposed model will be obtained and recorded to validate the model.

Figure 4 demonstrate the identification of tumor using DCNN optimization with skull eliminations.

Figure 4

The segmentation results

Table 1

Parameter Values of PSO, GA, GWO and WOA

PSO parameters for proposed algorithm	GA parameters for proposed algorithm	GWO parameters for proposed algorithm	WOA parameters for proposed algorithm
1. Number of Particles=16	1. Number of population=16	1. Number of wolves=16	1. Number of wolves=16
2. Maximum Number of Iterations=100	2. Maximum Number of Iterations=100	2. Maximum number of iterations=100	2. Maximum number of iterations=100
3. Inertia Weight=1	3. Mutation Coefficient (alpha)=0.2	3. Alpha wolf initial value=inf.	3. Initial search agent value=inf.
4. Inertia Weight Damping Ratio=0.99	4. Elite count=Roulette wheel selection	4. Beta wolf initial value=inf.	4. Acceleration co-efficient=linearly decrease
6. Personal Learning Coefficient=1.5		5. Delta wolf initial value=inf.
8. Global Learning Coefficient=2.0		6. Acceleration co-efficient=linearly decrease

4 Experiment and Result Discussion

The research work has carried out by demonstrating the outcomes of the proposed model applied on brain MRI of BRATS 2015 database for performance analysis of the proposed method.

The experiments follow the steps required to train and segment the brain tumor with CNN optimization.

1. The result of CNN process is optimized by the output vector of CNN with whale optimization (WOA). Whale optimization serves to make the value of the loss function on CNN to be at minimum.

2. The outcome vector will be improved if the solution of current whale value has less error than the earlier vector output.

3. The whale optimization will continue till the convergence solution is found.

4. After the CNN training, the model will be tested using the testing data (Ground truth).

5. The outcomes of the CNN vector represent the accuracy of CNN to predict the actual value of brain tumor segmentation with contrast to the ground truth.

Here we have used the different parameters for different optimization technique for our simulations as illustrated the table below.

In this paper, we have proposed to apply optimization algorithm in deep convolution neural network to optimize its network parameters such as weights and biases. The proposed model is tested with MRI brain tumor image for segmentation with respect to the validation indexes such as specificity, sensitivity and accuracy.

We demonstrate the convergence behavior of the different optimization algorithms used in DCNN for reducing errors (MSE). During the process, the co-efficient of algorithm was decreased gradually with iterations and consequently the more linearly decreased algorithm was found efficient.

As shown in the Figure 12, the accuracy of the proposed model along with Whale optimization algorithm gives better result as compared to other optimization techniques like PSO, GA and GWO for CNN model based on BRATS 15 dataset. From the convergence curves of each individual model shown in Figures 5 to 8 and the comparison of the convergence characteristics of all 4 models in Figure 9, we observed that the DCNN algorithm model is converges faster by not getting trapped into local minima. To validate this model, the accuracy of the 4 models and ROC curves can also be plotted with the rate and false positive rate in Figures 11 and 10 respectively.

Figure 5

Convergence curve of CNN-GA

Figure 6

Convergence curve of CNN-PSO

Figure 7

Convergence curve of CNN-GWO

Figure 8

Convergence curve of CNN-WOA

Figure 9

Comparison result of Convergence curve of different proposed models

Figure 10

Accuracy of different proposed models i.e., CNN-GA, CNN-PSO, CNN-GWO and CNN-WOA

Figure 11

ROC curve between true positive rate and false positive rate between proposed algorithms

Figure 12

Comparison of different proposed algorithms

5 Conclusions and Future Work

Deep convolutional neural network extracts the relevant features automatically. In order to get optimized learning parameters of the network, swarm optimization-based algorithms are applied in CNN network to make more suitable and faster to learn from training samples. In this investigation we used the 4 optimization algorithms such as (i.e., WOA, GWO, GA, and PSO) to enhance the performance and reduce the loss function in DCNN training model. During testing, the accuracy record for GA, PSO, GWO and WOA guided DNN was 0.958, 0.960, 0.961, 0.980 and 0.960, 0.963, 0.963, 0.964 and 0.932, 0.957, 0.962, 0.974 in specificity, sensitivity, and accuracy respectively. These performance measures are presented in the form of bar chart in Figure 12 and it is clearly visible from the figure that WOA-DCNN model performs better than the other 3 optimizations based DCNN models.