An improved framework for brain tumor analysis using MRI based on YOLOv2 and convolutional neural network

Brain tumor is a group of anomalous cells. The brain is enclosed in a more rigid skull. The abnormal cell grows and initiates a tumor. Detection of tumor is a complicated task due to irregular tumor shape. The proposed technique contains four phases, which are lesion enhancement, feature extraction and selection for classification, localization, and segmentation. The magnetic resonance imaging (MRI) images are noisy due to certain factors, such as image acquisition, and fluctuation in magnetic field coil. Therefore, a homomorphic wavelet filer is used for noise reduction. Later, extracted features from inceptionv3 pre-trained model and informative features are selected using a non-dominated sorted genetic algorithm (NSGA). The optimized features are forwarded for classification after which tumor slices are passed to YOLOv2-inceptionv3 model designed for the localization of tumor region such that features are extracted from depth-concatenation (mixed-4) layer of inceptionv3 model and supplied to YOLOv2. The localized images are passed to McCulloch's Kapur entropy method to segment actual tumor region. Finally, the proposed technique is validated on three benchmark databases BRATS 2018, BRATS 2019, and BRATS 2020 for tumor detection. The proposed method achieved greater than 0.90 prediction scores in localization, segmentation and classification of brain lesions. Moreover, classification and segmentation outcomes are superior as compared to existing methods.


Introduction
The tumor is a mass of irregular cells called the primary brain tumor inside the brain. The common symptoms of brain tumors are headaches, seizures, difficulties in speech, vomiting, imbalance problem, sensation loss, changes in behavior, and personality [58]. In America, 700,000 persons are suffering from brain tumor, and expected to increase to more than 79,000 by the end of 2020. Among these, 25,000 may suffer from malignant and remaining from non-malignant tumor [15]. Glioma is a predominant form of brain tumor, broken into low-and high-grade brain tumors. such that high grade is more aggressive as compared to low grade [13]. MRI is utilized to examine anatomical body structure [20,32], which is widely used for the detection of brain tumors. An error-prone and more exhaustive activity is manual diagnosis of brain tumors using MRI. Therefore, automated approaches are used for anomalous detection which is helpful for accurate and fast detection [2,[8][9][10][42][43][44][45]. Nowadays, several researchers are focused on different imaging sequences of MRI to analyze the tumor region [9,61,66]. Several techniques are introduced in literature based on clustering [19,31,47] and super pixels [54] for brain tumor detection. Appropriate extraction of features and optimization is a difficult task i.e., [56], particle swarm optimization (PSO) [31], local binary patterns (LBP), and histogram features [1,59] are utilized for the classification of tumor. The existing approaches have failed for the detection of more than one small volume of tumor per MRI slices [29]. These methods detect tumors on only Flair imaging modality such that SVM has been utilized for classification that performed better on small data. Hence, there is still a need of improved techniques for tumors detection on different views, such as saggital, coronal, and axial from large-scale imaging data [5,14]. Keeping this in view, an improved approach is presented in this article for classification, localization, and segmentation of glioma lesions. The major article contribution is opted as follows: • The homomorphic wavelet filer is applied on input MRI images for noise removal and passed to the pre-trained inceptionv3 model for feature extraction, where optimum features are selected using NSGA. • After classification, infected region is localized based on YOLOv2-inceptionv3 model, where deep features are extracted using depth-concatenation (mixed-5) layer and passed to YOLOv2 model. • McCulloch's Kapur entropy is applied to localized images for 3D-segmentation of tumor region. The segmentation outcome is also validated with truth annotated images to confirm the method's effectiveness.
The remaining manuscript is divided in different sections i.e., related work is in "Related work", and proposed work with respected results are presented in "Proposed methodology" and "Results and discussion", respectively.

Related work
Extensive work has been done for brain tumor detection [11]. Enhancement is a more vital task for noise reduction that aids in the improvement of segmentation. Wavelet filter [50], median filter [7], Gaussian filter [52], PDDF filter, FNLM filter [49], and high-pass filter [7] are used in pre-processing step. Pereira et al. [41] applied CNN with 3 kernel sizes and obtained 0.88, 0.83, 0.77 dice scores of complete, enhance, and non-enhanced tumor regions, respectively. Sauwen et al. [48] proposed different methodologies to analyze tumor segmentation results [26]. Goswami and Bhaiya [6] presented a hybrid framework consisting of fuzzy logic and neural network for tumor detection and classification [51]. A semiautomatic method with spatial features is applied for tumor detection [24]. Different clustering approaches (K-means [8], PSO, MFKM) are used for the segmentation of tumor [60]. Watershed is utilized with GLCM for features extraction and supplied to SVM [53] for multi-fractals classification with a higher precision rate. The transfer learning models are widely utilized to classify the tumor region, such as Alex-net, Google Net, and VGG-16. Two different types of neural networks are trained on augmented input images for brain lesions classification [52]. The pre-trained AlexNet has been utilized for glioma detection for the prediction of patient's survival rate [53]. CNN model is trained on brain imaging data and classified input data into five classes, such as multiform glioma, astrocytoma, shapeless tumor, normal brain tissues, and oligodendroglioma [6]. M-net segmentation model has been utilized for features extraction and fed into the pre-trained VGG-16 for the classification of three different types of the tumor [63]. Fuzzy-c-means has been applied for segmentation followed by DWT features extraction and suitable features selection by PCA for classification [35]. Capsule Networks (CapsNets) has been utilized [3]. 3-D CNN architecture has been utilized for glioma classification into different grades, such as low and high [23]. 2-D-CNN has been used for increasing the precision rate of glioma classification [21,22]. Deep CNN network has been applied for glioma classification. 3D-Unetwork has been used for glioma detection in which average global pooling layer is used for features mapping followed through 1 × 1 cascade convolutional work as FC layer [7]. A CNN model is utilized for deep features extraction and informative features selection using GA for glioma classification [12]. While comprehensive tumor detection and classification work has been performed, but still accurate tumor detection is a challenging task and has room for improvement. Therefore, this research work provides an improved approach for classification, localization, and segmentation of brain tumor.

Proposed methodology
The proposed method has four primary steps: (1) enhancement, (2) classification, (3) localization, and 4) segmentation as illustrated in Fig. 1 such that input images are enhanced using homomorphic wavelet filer and classified using extracted deep features from inceptionv3. The classified images are localized through the proposed YOLOv2-inceptionv3 and segmented based on Kapur entropy.

Noise elimination using homomorphic wavelet filter
The images acquired from MRI protocol having adversative situations might be contaminated due to noise that degrades the disease detection rate. Several filters are presented for noise removal. These filters depend on noise type included in the images. Wavelet transform is used to represent the images into frequency domain. In this process, image decomposition is performed to process the image into high-high (HH), low-high (LH), and high-low (HL) bands. This research investigates a homomorphic wavelet 1 3 filter decomposition to eliminate speckle noise that is mathematically expressed as follows: The noise removal process using a homomorphic filter with wavelet decomposition is visualized in Fig. 2 such that image is decomposed into 04 bands HL, LH, HH, and LH-HH. The HH band improves the image quality as compared to other bands like HL, LH, and LH-HH. Thus, for further processing, HH band is utilized to perform accurate segmentation.

Extracted deep features using pre-trained inceptionv3 architecture
Deep learning is widely utilized in artificial intelligence applications, such as speech recognition and computer vision. However, with more interest in the area of deep learning, classification into corresponding categories is a major problem. This problem might be solved through transfer learning because accurate models and architecture are built in in a time-saving manner. In this process, learning is performed through already learned patterns to solve different problems instead of using features learning from scratch. Transfer learning uses pre-trained models that are learned on huge amount of data for problem-solving. Thus, this work utilizes an inceptionv3 pre-trained transfer learning model [55] for features learning which consists of 01 image, 094 Convolutional (Conv), 094 batch-normalization (bn), 094 ReLU, 14 max-pooling, 015 depth concatenation, fully connected layers, and softmax with cross-entropy function. The features are extracted from fully connected layers named as prediction and further passed to NSGA [18] for improved features selection as displayed in Fig. 3.

Features Selection
A deep feature vector (1 × 1000) is obtained using incep-tionv3 network. The features of engineering are performed to select optimum feature vector by applying NSGA II. The parameters of NSGA as are discussed in Table 1.

Localization using YOLOv2-inceptionv3 model
YOLOv2-inceptionv3 model with 174 layers is proposed to localize tumor region such that there is 165 layers of inceptionv3 with 01 input, 50 Conv, 50 bn, 50 activations ReLU, 06 mixed (depth concatenation) 03 max-pooling, 05 average pooling, and 09 layers of tinyYOLOv2 [46] model. The optimum hyper-parameters are discussed in Table 2.   The proposed model more accurately localizes tumor region as illustrated in Fig. 4.
YOLOv2 model optimized MSE loss among predicted bounding and ground truth boxes. The model training is performed on three different types of losses, such as localization, confidence, and classification. Among the expected and ground truth boxes, localization loss computes error using location, estimated box size, and ground truth. The confidence loss is utilized to compute objectiveness error with detected object in jth bounded box of grid i cell. The classification loss is used to calculate probability across each class of grid cell i. The mathematical formulation of these parameters is expressed as: Here, s represents grid cell, p denotes probability, w1, w2,w3 and w4 show weights, gc presents grid cell, x i ,ŷ i denotes center of bounding box, (x i , y i ) shows center of ground truth. (width i , height i ) signifies width and height of bounding box and (ŵidth ii ,ĥeight) presents width and height of ground truth. (2)

Lesion segmentation
A key challenge in medical images is variability in medical data. In human anatomy, variations occur in different modalities including X-ray, MRI, CT, and PET, etc. The segmentation region is used to analyze the disease severity levels. In the proposed method, McCulloch's Kapur entropy method [28] is utilized for tumor segmentation. In this method, probability of intensity values distribution is measured from the foreground and background regions after which entropy is calculated separately from both regions. The optimum value of threshold is applied to increase the sum of their entropies. The Kapur entropy is mathematically expressed as: Here (3)

Results and discussion
The method is evaluated on BRATS series including 2018, 2019, and 2020 [16,30,33]. BRATS 2018 contains 266 MRI patients with 191 high and 75 low glioma grade, BRATS 2019 composes of 285 patients, and BRATS 2020 has 335 patients such that each patient contains 155 slices. The detailed description of benchmark databases is illustrated in Fig. 6. The 0.5 hold-out validation approach is utilized for tumor slices classification, where half data are used for training and remaining for validation. The summary of classified images is given in Table 3.

3
The proposed work is evaluated on experiments implemented on MATLAB 2020-Ra toolbox with 2070 Nvidia Graphic Card and Gamming Laptop G5-5500 to validate the enhancement method, classification approach, localization technique, and segmentation method, respectively.

Experiment#1
In this experiment, the enhancement technique is evaluated in terms of different performance metrics, such as SSIM, MSE, and PSNR. The enhancements results are mentioned in Table 4 as well as visually presented in Fig. 7.
In Fig. 7, quantitative results are computed in terms of MSE, SNR, and PSNR using four bands, such as HL, LH, HH, and LH-HH. In this procedure, 80. 46 Table 4.
The results in Table 4 depict that proposed method att ained maximum 89.4096815346192 PSNR,    78.0342291390692 SNR, and 0.000369981932768593 MSE. The overall performance is represented in Fig. 8.

Experiment#2
In experiment#2, tumor predictions are done on 0.5 hold-out validation that is mentioned in Tables 5, 6, 7, 8. The method classified brain images (normal (0) and abnormal (1)) as shown in confusion matrices in Fig. 9. Figure 10 shows ROC on BRATS datasets with maximum 1.00 AUC and minimum 0.98 AUC.
In terms of performance metrics, BRATS 2018 obtained 0.0000 FNR while BRATS 2020 achieved 0.0138 FNR. Table 6 shows analysis of applying different classifiers to final features vector, where DT achieves 95% ACC, 97% SP, 88% SE, 85% PPV, 0.0300 FPR, and 0.1132 FNR. On discriminant analysis, quadratic kernel obtains highest results in comparison with linear kernel, such as 98% ACC on quadratic and 97% on linear kernel of LDA. On SVM, quadratic kernel attains 93% PPV and linear kernel shows 92% PPV.
The results in Table 7 show that DT achieves 94% ACC, while discriminant analysis shows 98% ACC using quadratic and 97% ACC using linear kernel. In geometrical family, SVM achieves 98% ACC on linear and 99% ACC on quadratic and cubic kernels. Tables 5, 6, 7, 8, SVM (cubic kernel) achieves maximum 0.9891 ACC whereas minimum 0.9563 ACC is obtained using DT on BRATS 2018. Likewise, on BRATS 2019, SVM (cubic kernel) attains maximum 0.9970 ACC and minimum 0.9421 ACC is obtained using DT. On BRATS 2020, SVM (quadratic kernel) shows maximum 0.9939 ACC while minimum 0.9329 ACC is attained with DT. Finally, it is observed that SVM performs better than other classifiers. Proposed method results comparison is stated in Table 9. Table 9 shows the results comparison with existing work, such as [4,37,65,67], such that 94% SE and 95% SE are obtained on BRATS 2018 while 96% SE is attained on BRATS 2019 datasets, respectively. However, SE of 100% and 99% are shown on BRATS 2018 and BRATS 2019 datasets, respectively, using proposed method.

Experiment#3
In this experiment, YOLOv2-inceptionv3 model is validated on performance metrics, such as mAP and IoU, as shown in Table 10 such that proposed method achieved mAP of 0.98, 0.99 and 1.00 on BRATS 2018, 2019 and 2020, respectively. The recommended approach localizes tumor region with highest confidence scores presented in the Fig. 11.

Experiment#4
In this experiment, localized images are segmented to analyze actual infected region more precisiely. The mathmatical formulation of segmentation measures, such as dice and jaccard index, is defined as: In this experiment, localized images are segmented to analyze the actual infected region more precisiely. The numerical computed results are also discussed in Table 11.
From the results in Table 11, it is observed that on HGG glioma, maximum 1.00 (dice, Jaccard index) and minimum 0.99, 0.98 (dice, Jaccard index) are obtained. On LGG, maximum 1.00 and minimum 0.99 (dice, Jaccard index) are achieved, respectively. The average segmentation outcomes on BRATS series are listed in Table 12. Table 12 shows that proposed framework achieved dice of 0.98, 0.96 and 0.97 on BRATS 2018, 2019 and 2020 datasets. The segmentation results on HGG and LGG are visualized in Figs. 12, 13, 14. The results comparison is given in Table 13. The proposed segmented results are compared with eight recent published works, such as [27,34,36,38,40,57,62]. The existing methods achieved maximum 0.82 dice score on 2018 BRATS, 0.89 dice score on 2019 BRATS and 0.84 dice score on BRATS 2020 datasets. In comparison with existing methods, presented framework achieved 0.98, 0.96 and 0.97 scores on BRATS 2018, 2019 and 2020 databases, respectively.    Kapur entropy method. The experimental results conclude that proposed approach achieved competitive results than the recent published work. The improved hybrid approach can be utilized in real-time applications to diagnose brain tumor at a premature stage. This research will be further expanded in future for the study of brain tumors using algorithms of quantum computation.
Author contribution All authors contributed equally to this research.
Funding There is no funding received for this research.

Declarations
Conflict of interest All authors declared that there is no conflict of interest.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat iveco mmons .org/licen ses/by/4.0/.   [36] 2019 0.89 [39] 0.59 [40] 0.82 Proposed Method 0.96 [38] 2020 0.80 [57] 0.80 [34] 0.84 [27] 0.81 Proposed Method 0.97