Detection of meningioma tumor images using Modified Empirical Mode Decomposition (MEMD) and convolutional neural networks

Abstract

The meningioma brain tumor detection is more important than the other tumor detection such as Glioma and Glioblastoma, due to its high severity level. The tumor pixel density of meningioma tumor is high and it leads to sudden death if it is not detected timely. The meningioma images are detected using Modified Empirical Mode Decomposition- Convolutional Neural Networks (MEMD-CNN) classification approach. This method has the following stages data augmentation, spatial-frequency transformation, feature computations, classifications and segmentation. The brain image samples are increased using data augmentation process for improving the meningioma detection rate. The data augmented images are spatially transformed into frequency format using MEMD transformation method. Then, the external empirical mode features are computed from this transformed image and they are fed into CNN architecture to classify the source brain image into either meningioma or non-meningioma. The pixels belonging tumor category are segmented using morphological opening-closing functions. The meningioma detection system obtains 99.4% of Meningioma Classification Rate (MCR) and 99.3% of Non-Meningioma Classification Rate (NMCR) on the meningioma and non-meningioma images. This MEMD-CNN technique for meningioma identification attains 98.93% of SET, 99.13% of SPT, 99.18% of MSA, 99.14% of PR and 99.13% of FS. From the statistical comparative analysis of the proposed MEMD-CNN system with other conventional detection systems, the proposed method provides optimum tumor segmentation results.

Keywords

Meningioma tumor transformation features classification rate

1 Introductions

The fast growing cells in human brain leads to the development of the tumor. These affected cells spread to the other regions of the brain which leads to the death. The tumor cells can be categorized into Low Density Profile (LDP) and High Density Profile (HDP) cases according to the spreading area of the abnormality portions in human brain. The slow development of tumor cells in brain is called as LDP and its severity level is low. The fast development of tumor cells in brain is called as HDP and its severity level is high. LDP tumors are otherwise called as primary type tumors and it does not lead to sudden death. HDP tumors are otherwise called as secondary type tumors and it leads to sudden death. Hence, the detection of HDP is more important than the detection of LDP [1 –4]. The tumor regions in brain can be scanned through different scanning procedures such as Computer Tomography (CT) and Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET). The brain tumors are clearly visible in MRI and PET scanning procedures than the CT. The radiation level used in PET for scanning the brain regions is significantly higher than the radiation levels of the MRI. Therefore, MRI scanning procedure is mostly preferred for scanning the brain regions with high density pixel profile [5, 6]. Magnetic Resonance Imaging is widely used, due to its high transparency properties in the soft tissues of the brain and is a non-ionizing, non-radioactive substance that will not harm human organs. Combined with medical knowledge and clinical experience, a knowledgeable physician can determine tumor size, shape, anatomical structure and other pathological features of brain tumors that help promote appropriate treatment for patients.

Glioma, Meningioma and Glioblastoma tumors are the category of the HDP cases and they are depicted in Fig. 1 (a–c) respectively. The pixel density of tumor region in Meningioma case is comparatively higher than the pixel density of tumor regions in both Glioma and Glioblastoma cases. Among these tumor types, Meningioma is mostly occurred in all age groups with different stages of the severity [7, 8]. Meningiomas are neoplasm that are produced by meningothelial cells with an incidence of 0.9% in routine brain Magnetic Resonance Imaging diagnosis. However, nearly 30% of the primary intracranial lesions are considered as meningiomas. According to the World Health Organization (WHO), the diagnosed lesion can be graded as benign (grade I), atypical (grade II) or anaplastic (grade III). The histological grading predicts the lesions in behavioural model and prognosis of meningiomas. Grade II and III type of meningiomas are proved to be more invasive and aggressive. By evaluating histological Criteria, the Grade I & II show atomic atypia, Grade III secondary cell segmentation density, atomic atypia, mitotic action, Grade IV atomic atypia, mitotic activity, vascular burgeoning or corruption.

Fig. 1

(a) Glioma (b) Meningioma (c) Glioblastoma.

The conventional meningioma brain tumor detection requires high number of radiologist to identify the tumor regions in high population countries like India. Also, the time consumption is high for detecting the tumor regions. The manual tumor detection process also exhibits the error during the tumor region segmentation. Hence, there is a scope for developing the automatic computer aided procedures for identifying the meningioma tumor images to prevent earlier death. In this article, the deep learning soft computing method is used to screen the meningioma tumor images with the integration of EMD transformation approach.

2 Literature survey

Muhammad Arif et al. (2022) used Biologically Inspired Orthogonal Wavelet Transform (BIOWT) for decomposing the brain image into various number of sub band images. These decomposed sub bands were further used for the classification model incorporating CNN. This approach was evaluated on set of meningioma brain MRI images and the authors obtained 96.89% of SET, 97.92% of SPT, 95.89% of MSA, 98.28% of PR and 97.49% of FS on meningioma brain images. Arkapravo Chattopadhyay et al. (2022) combined machine and deep learning methods to locate the tumor Region of Pixels (RoP) in tumor images. The authors used conventional Support Vector Machine (SVM) algorithm for extracting the internal features shape, size and location and these computed SVM incorporating metrics were used for classification. This methodology was evaluated on set of meningioma brain MRI images and the authors obtained 97.48% of SET, 95.92% of SPT, 96.39% of MSA, 97.13% of PR and 97.67% of FS on meningioma brain images.

Irmak et al. (2021) used deep Convolutional neural network architecture to identify the tumor region of pixels in meningioma brain images. The learning rate was optimized using activation function in each internal layers of designed Convolutional layer. The number of each internal layers were specified based on the number of internal features required for the effective classification of images. The authors obtained 97.10% of SET, 97.82% of SPT, 96.86% of MSA, 97.04% of PR and 97.18% of FS on meningioma brain images. Guoli Song et al. (2021) used extended membership filter which was integrated with the back propagation neural network classification approach for the detection of tumor brain images from the non-tumor brain images. The membership functions was initially set based on the Grey Level Co-occurrence Matrix (GLCM) feature matrix and they were dynamically changed based on the number of extracted internal features. The authors obtained 95.98% of SET, 96.89% of SPT, 97.17% of MSA, 96.38% of PR and 97.94% of FS on meningioma brain images.

Balakumaresan Ragupathy et al. (2020) used machine learning algorithm CoActive Adaptive Neuro Fuzzy Inference System (CANFIS) classifier and U-Net CNN classifier for the detection of meningioma brain images on Nanfang university dataset. The meningioma tumor images were classified using various non-linear set of features with Non-Sub sampled Contourlet Transform (NSCT) transform. The authors used U-Net CNN architecture for both detection and segmentation of tumor regions in meningioma brain MRI images. The authors obtained 96.26% of SET, 96.64% of SPT, 97.18% of MSA, 96.26% of PR and 97.38% of FS on meningioma brain images.

Jasmine Hephzipaha et al. (2020) derived heuristic set of features from both meningioma and non-meningioma brain images and these features were learned through optimization algorithm. The optimized and learned features were classified using Adaptive Neuro Fuzzy Inference System (ANFIS) classifier to classify these features which were belonging to meningioma or non-meningioma brain images. Then, the authors used morphological segmentation approach to segment the tumor regions. The authors obtained 96.1% of SET, 96.3% of SPT, 97.2% of MSA, 97.8% of PR and 97.1% of FS on meningioma brain images. Selvapandian et al. (2018) proposed Gradient Boosting Machine Learning (GBML) and classification approach for the classification of meningioma brain MRI images. The intensity features and Gray Level Run Length Matrix features (GLRLM) were computed from the source meningioma and healthy brain images. These regression set of features were classified by GBML classification method. This methodology was evaluated on set of meningioma brain MRI images and the authors obtained 97.28% of SET, 96.96% of SPT, 97.79% of MSA, 97.38% of PR and 97.49% of FS on meningioma brain images.

In this work, the advantages of conventional methodologies Irmak et al. (2021) and Guoli Song et al. (2021) were combined to develop a proposed methodology. The researchers in past decades in the field of brain tumor detection used either feature combining methods or deep learning methods. In this paper, both the methods are combined to develop a novel proposed meningioma brain tumor detection system to eliminate such limitations in conventional meningioma brain tumor detection methodologies.

The main objectives of this paper are stated below.

To develop a fully computer aided systematic meningioma brain tumor detection and segmentation method.

To propose MEMD-CNN classification approach to classify and detect the tumor regions in the meningioma brain images.

To validate the performance study of the proposed meningioma detection system with previous studies.

3 Proposed work

3.1 Dataset details

The images from BRAINWEB [15] dataset and Nanfang University dataset [16] are accessed for evaluating the performance of the meningioma tumor detection process. The BRAINWEB dataset is used in this paper to verify the effectiveness of the proposed methodology stated in this work. The image pixel resolution is about 1024×1024 with 8-bit quantization. The 400 meningioma brain images are split into training dataset (200 images) and testing dataset (200 images). Also, the 800 non-meningioma brain images are split into training dataset (400 images) and testing dataset (400 images). In this paper, 571 meningioma brain images and 750 non-meningioma brain images are obtained from the Nanfang University hospital dataset [15]. The image pixel resolution is about 2024×2024 with 8-bit quantization. The BRAINWEB dataset consist of low resolution brain images and Nanfang University dataset consist of high resolution brain images. Most of the researchers in previous studies applied and tested their proposed approaches on high resolution dataset images only. These conventional methods failed when it is applied and tested on low resolution dataset images. Hence, this paper uses both low and high resolution imaging datasets for evaluating the performance of the proposed methods stated in this manuscript.

3.2 Methods

In this article, the meningioma brain images are detected using MEMD-CNN classification approach. This proposed method has the following stages data augmentation, spatial-frequency transformation, feature computations, classifications and segmentation. The brain image samples are increased using data augmentation process for improving the meningioma detection rate. The data augmented images are spatially transformed into frequency format using MEMD transformation method. Then, the external empirical mode features are computed from this transformed image and they are fed into CNN architecture to classify the source brain image into either meningioma or non-meningioma. The tumor pixels are segmented in meningioma brain image using morphological opening-closing functions. The entire procedure for meningioma detection system is illustrated in Fig. 2(a) and 2(b) respectively.

Fig. 2

Meningioma brain image classification system (a) Learning mode (b) Testing mode.

3.3 Data augmentation

The image count in dataset is improved by increasing the number of training samples for training the classification architecture. In this article, 571 meningioma with 750 non-meningioma images are obtained from the Nanfang University hospital dataset [16]. The image pixel resolution is about 1024×1024 with 8-bit quantization. These 571 meningioma brain images are split into training dataset (271 images) and testing dataset (200 images). Also, the 750 non-meningioma brain images are split into training dataset (500 images) and testing dataset (250 images). Total training images used for training the CNN classifier is about 771 and it is not adequate for obtaining the high classification rate. Hence, data augmentation procedure is used in this article to improve the training sample images. In this article, left pixel flip, right pixel flip, left pixel rotate and right pixel rotate data augmentation functions are used. Hence, the total number of 771 images are data augmented into 3084 images and these images are integrated with non-data augmented images (771 images) to obtain 3855 data augmented images.

Total training images from BRAINWEB dataset used for training the CNN classifier is about 600 and it is not adequate for obtaining the high classification rate. Hence, data augmentation procedure is used in this article to improve the training sample images. In this article, left pixel flip, right pixel flip, left pixel rotate and right pixel rotate data augmentation functions are used. Hence, the total number of 600 images are data augmented into 2400 images and these images are integrated with non-data augmented images (600 images) to obtain 3000 data augmented images.

3.4 Modified EMD transformation model (MEMD)

The geometrical transform transforms the pixel having spatial mode to the frequency mode of pixels. In conventional methods, Discrete Wavelet Transform (DWT), Gabor and Contourlet transforms are available for pixel transformations of the image. There is spatial loss of pixels occurred during the transformation of pixels, which degrades the performance efficiency. Hence, there is a need for less spatial loss during transformation of pixels in brain image. In order to achieve this, Empirical Mode Decomposition (EMD) transformation model is proposed to mitigate the drawbacks of the existing transformation models. Pixel localization and directionality properties are the key advantage of this transform to reduce the spatial errors. In general, EMD transform is used for decomposition of signals. In this article, the MEMD transform is proposed for image decomposition by modifying the decomposition procedure of the EMD transform.

The procedure for the MEMD model is explained in the following Algorithm.

Algorithm: MEMD

Input: Source image;

Output: MEMD transformed image;

Start;

Step 1:

The pixels in the source brain image are stored in a matrix, which is called as Mode Matrix (MM), where the rows and columns of this matrix is similar to the rows and columns of the source image.

Step 2:

Determine the maximum pixel value in MM using Equation (1), $P_{\max} = Maximum (MM)$ (1) Where, MM is the Mode Matrix.

Step 3:

Determine the minimum pixel value in MM using Equation (2), $P_{\min} = Minimum (MM)$ (2)

Step 4:

Find the Absolute Averages (AA) between P_max and P_min, using the following equations. ${AA}_{1} = \frac{1}{2} (P_{\max} + P_{\min})$ (3) ${AA}_{2} = \frac{1}{2} (P_{\max} - P_{\min})$ (4) ${AA}_{3} = Max (P_{\max} + P_{\min})$ (5) ${AA}_{4} = Min (P_{\max} + P_{\min})$ (6) ${AA}_{5} = Max (P_{\max} - P_{\min})$ (7) ${AA}_{6} = Min (P_{\max} - P_{\min})$ (8)

Step 5:

Find the Residual sub bands (R_i) using the computed AA with respect to the following equations. $R 1 = MM - {AA}_{1}$ (9) $R 2 = MM - {AA}_{2}$ (10) $R 3 = MM - {AA}_{3}$ (11) $R 4 = MM - {AA}_{4}$ (12) $R 5 = MM - {AA}_{5}$ (13) $R 6 = MM - {AA}_{6}$ (14)

The computed residual sub bands size is equal to the source image matrix (MM) size.

Step 6:

The final decomposed residual sub band image is constructed by considering the maximum pixel value of each pixel position in all six residual sub band images, using the following equation. $R_{combined} = Maximum (R 1, R 2, R 3, R 4, R 5, R 6)$ (15)

End;

The MEMD method decomposes the data augmented brain image into six sub band images and these six sub band images are combined into final decomposed image by selecting the maximum value in each sub bands illustrated in the proposed algorithm.

3.5 External feature derivation

The features correlate the relationship between tumor pixel and non-tumor pixel in the image. In this paper, both internal and external features are computed from the source brain image and they are used for tumor and non-tumor pixel classifications. The external features are computed from the MEMD transformed image and the internal features are computed from the internal layers of the deep learning CNN architecture. In this section, external features are derived from the MEMD transformed image. $Mean (M) = \sum_{i = 1}^{M} \sum_{j = 1}^{N} {EMD}_{ij}$ (16)

Where, EMD is the decomposed MEMD transformed image, M and N represents the number of rows and columns of the MEMD transformed image. $Variance (V) = \sum_{i = 1}^{M} \sum_{j = 1}^{N} {({EMD}_{ij} - M)}^{2}$ (17)

The following external features are computed using the Mean and Variance features.

$\begin{matrix} Empirical Mean Features (EMF) = \frac{1}{M \times N} \\ \sum_{i = 1}^{M} \sum_{j = 1}^{N} (\frac{({EMD}_{ij} - M) \times ({EMD}_{ij} + M)}{M}) \end{matrix}$ (18)

$\begin{matrix} Empirical Variance Features (EVF) = \frac{1}{M \times N} \\ \sum_{i = 1}^{M} \sum_{j = 1}^{N} (\frac{({EMD}_{ij} - V) \times ({EMD}_{ij} + V)}{V}) \end{matrix}$ (19) $\begin{matrix} Emirical Mean Energy Features (EMEF) = \frac{1}{M \times N} \\ \sum_{i = 1}^{M} \sum_{= 1}^{N} (\frac{{({EMD}_{ij} - M)}^{2} \times {({EMD}_{ij} + M)}^{2}}{M}) \end{matrix}$ (20) $\begin{matrix} Emirical Variance Energy Features (EVEF) = \frac{1}{M \times N} \\ \sum_{i = 1}^{M} \sum_{= 1}^{N} (\frac{{({EMD}_{ij} - V)}^{2} \times {({EMD}_{ij} + V)}^{2}}{V}) \end{matrix}$ (21)

The computed external features are stored in a Matrix Vector (MV) and it is fed into the CNN classifier.

3.6 Classifications

It is the process of classifying the computed MV features into either Meningioma or Non-Meningioma case. The process of classification can be done by either machine or deep learning classifiers. The machine learning classifiers such as Neural Networks (NN) and Adaptive Neuro Fuzzy Inference System (ANFIS) requires more training samples in the mode of the classification process for achieving high classification rate. This increased the computation time as the main limitation of the conventional machine learning classifier. In order to overcome such limitation, this paper uses deep learning architecture which classifies the brain image into either Meningioma or Non-meningioma with less detection time.

In conventional AlexNet CNN architecture, there are five numbers of Convolutional Layers (CL) and five numbers of down sampling layers and three numbers of Fully Connected Neural Network (FCNN) layer. The proposed AlexNet CNN architecture is designed with three numbers of CL, two numbers of down sampling layers and two numbers of FCNN, as illustrated in Fig. 3. The first CL is constructed with 256 numbers of filters with the kernel size of 5×5, the second CL is constructed with 512 numbers of filters with the kernel size of 5×5 and the third CL is constructed with 256 numbers of filters with the kernel size of 7×7. The number of filters and its kernel size is chosen after several iterations in order to obtain high classification rate. The MV is fed into the first CL which is convolved with the kernel of each filters and produces the internal features. These internal features are again fed into second CL for producing the internal features. The size of these internal features from CL-2 is high which increases the detection time. In order to reduce the detection time of the classification process, down sampling layer is used which reduces the size of the internal features generated through CL. The down sampling layer is designed with 3×3 Max-pooling functions instead of Average pooling function to reduce the down sampling losses during pooling process. The down sampled sequences are now passed through the CL-3, which also produces the internal features. The size of these internal features is reduced using down sampling layer 2. The down sampled sequences are finally fed into FCNN, which is designed with layer 1 (4096 neurons) and layer 2 (2 neurons). Activation of first neuron in layer 2 represents the meningioma case and the activation of second neuron in layer 2 represents the non-meningioma case.

Fig. 3

CNN architecture for meningioma image classifications.

Table 1 shows the CNN architecture layer specifications in both conventional and proposed AlexNet classifiers.

Table 1

CNN architecture layer specifications

Layer specifications	Conventional AlexNet (Prilianti et al. 2019)	Proposed AlexNet
Convolutional layer-1	96 filters with 11×11 kernel size	256 filters with 5×5 kernel size
Convolutional layer-2	256	512 filters with 5×5 kernel size
Convolutional layer-3	384	256 filters with 7×7 kernel size
Convolutional layer-4	384	–
Convolutional layer-5	256	–
FCNN layers	Three FCNN layers: 4096 neurons in layer 1, 4096 neurons in layer 2, 100 neurons in layer 3.	Two FCNN layers: 4096 neurons in layer 1, 2 neurons in layer 2

Finally, morphological segmentation approach (eroded functional image is subtracted from dilated functional image) is used to segment the tumor regions in the classified meningioma brain image. Figure 4(a) is the classified meningioma brain image and Fig. 4(b) is the tumor region segmented meningioma brain image respectively.

Fig. 4

(a) Meningioma brain image (b) Tumor region segmented image.

4 Results and discussions

The meningioma brain tumor detection method stated in this paper is experimentally evaluated and analyzed using MATLAB R2020 version with Core i5- 3.8 GHz dual core processor and 12 GB internal RAM as the hardware specifications. This meningioma detection method is experimentally analyzed in term of Meningioma Classification Rate (MCR) and Non-Meningioma Classification Rate (NMCR). MCR is computed between meningioma image count and the total meningioma image count and measured in percentage. NMCR is computed between the detected non-meningioma image count and the total non-meningioma image count and measured in percentage.

In this article, 571 meningioma and 750 non-meningioma brain images are used from the open access dataset. The MEMD-CNN methodology stated in this article achieves 99.4% of MCR by classifying 568 meningioma images correctly over 571 meningioma images. The MEMD-CNN methodology also achieves 99.3% of NMCR by classifying 745 non-meningioma images correctly over 750 non-meningioma images. Therefore, the average Classification Rate (CR) of the MEMD-CNN method is about 99.3%.

Table 2 is the experimental analysis of meningioma and non-meningioma detection system with respect to multi resolution transforms on Nanfang university dataset images. The proposed meningioma detection system is tested by implementing conventional EMD transform and proposed MEMD transform and the experimental results are analyzed with respect to MCR and NMCR, as illustrated in Table 2. The meningioma detection system using conventional EMD transform obtains 92.9% of MCR and 97.6% of NMCR, whereas the meningioma detection system using MEMD transform obtains 99.4% of MCR and 99.3% of NMCR on the meningioma and non-meningioma images available in open access dataset.

Table 2
Experimental analysis of meningioma and non-meningioma detection system with respect to multi resolution transforms on Nanfang university dataset images

Transformation model Meningioma images tested count Non-meningioma images tested count Correctly classified meningioma image count Correctly classified non-meningioma image count MCR (%) NMCR (%)

EMD 571 750 531 732 92.9 97.6

MEMD 571 750 568 745 99.4 99.3

Transformation model	Meningioma images tested count	Non-meningioma images tested count	Correctly classified meningioma image count	Correctly classified non-meningioma image count	MCR (%)	NMCR (%)
EMD	571	750	531	732	92.9	97.6
MEMD	571	750	568	745	99.4	99.3

Table 3 is the experimental analysis of meningioma and non-meningioma detection system with respect to multi resolution transforms. The proposed meningioma detection system is tested by implementing conventional EMD transform and proposed MEMD transform and the experimental results are analyzed with respect to MCR and NMCR, as illustrated in Table 3. The meningioma detection system using conventional EMD transform obtains 91.5% of MCR and 88% of NMCR, whereas the meningioma detection system using MEMD transform obtains 99% of MCR and 99% of NMCR on the meningioma and non-meningioma images available in open access dataset.

Table 3

Experimental analysis of meningioma and non-meningioma detection system with respect to multi resolution transforms on BRAINWEB dataset images

Transformation model	Meningioma images tested count	Non-meningioma images tested count	Correctly classified meningioma image count	Correctly classified non-meningioma image count	MCR (%)	NMCR (%)
EMD	200	200	183	176	91.5	88
MEMD	200	200	198	198	99	99

The external features which are derived from the MEMD decomposed image plays a significant role in the classification results. Table 4 shows the experimental analysis of meningioma and non-meningioma detection system with respect to the external features on Nanfang university dataset images. Table 5 shows the experimental analysis of meningioma and non-meningioma detection system with respect to the external features on BRAINWEB dataset images.

Table 4

Experimental analysis of meningioma and non-meningioma detection system with respect to the external features on Nanfang university dataset images

External features combinations	MCR (%)	NMCR (%)
M+V	65.1	68.3
M+EMF	68.2	71.7
M+EMF+EVF	72.9	73.2
M+EMF+EVF+EMEF	73.2	75.6
M+V+EMF+EVF	74.3	75.9
M+EMF+EVF+EMEF+EVEF	76.9	79.3
M+V+EMF+EVF+EMEF+EVEF	99.4	99.3

Table 5

Experimental analysis of meningioma and non-meningioma detection system with respect to the external features on BRAINWEB dataset images

External features combinations	MCR (%)	NMCR (%)
M+V	65.3	66.9
M+EMF	68.9	69.3
M+EMF+EVF	75.6	76.8
M+EMF+EVF+EMEF	77.9	80.5
M+V+EMF+EVF	77.5	80.4
M+EMF+EVF+EMEF+EVEF	81.5	82.6
M+V+EMF+EVF+EMEF+EVEF	99	99

Further, the EMD-CNN meningioma detection method is experimentally analyzed with respect to the following confusion metrics. The confusion metrics is formed by computing actual values (tumor) and the predicted values (tested) in terms of positive and negative rate and it is defined in the following Table 6.

Table 6

Confusion matrix

		Actual values (tumor case)
		Positive	Negative
Predicted values	Positive	S _tp	S _fp
(tested)	Negative	S _fn	S _tn

The relation between S_tp, S_fp, S_fn and S_tn parameters are defined confusion matrix. The actual tumor-positive case and the predicted tumor-positive case represent S_tp. The actual tumor-negative case and the predicted tumor-positive case represent S_fp. The actual tumor-positive case and the predicted tumor-negative case represent S_fn. The actual tumor-negative case and the predicted tumor-negative case represent S_tn.

These values in Table 6 are used to find the performance evaluation metrics sensitivity, specificity, accuracy, precision and F1-score, using the Equations (22–26).

From the confusion matrix (Table 6), the following metrics are derived to evaluate the performance of MEMD-CNN methodology for meningioma detection system. $Sensitivity (SET) = \frac{S_{tp}}{S_{tp} + S_{fn}}$ (22) $Specificity (SPT) = \frac{S_{tn}}{S_{tn} + S_{fp}}$ (23)

$\begin{matrix} Meningioma Segmentation Accuracy (MSA) = \\ \frac{S_{tp} + S_{tn}}{S_{tp} + S_{tn} + S_{fp} + S_{fn}} \end{matrix}$ (24) $Precision (PR) = \frac{S_{tp}}{S_{tp} + S_{fp}}$ (25) $F 1 - Score (FS) = \frac{2 * S_{tp}}{2 * S_{tp} + S_{fp} + S_{fn}}$ (26)

Whereas, S_tp is number of true positive pixels with respect to actual case (positive) and predicted case (positive), S_tn is number of true negative pixels with respect to actual case (negative) and predicted case (negative), S_fp is number of false negative pixels with respect to actual case (negative) and predicted case (positive), S_fn is number of false positive pixels with respect to actual case (positive) and predicted case (negative).

Table 7 is the experimental analysis of MEMD-CNN for meningioma detection system. The proposed MEMD-CNN method for meningioma classification approach attains 98.93% of SET, 99.13% of SPT, 99.18% of MSA, 99.14% of PR and 99.13% of FS.

Table 7

Experimental analysis of MEMD-CNN for meningioma detection system

Meningioma	Experimental metrics in %
image sequences	SET	SPT	MSA	PR	FS
M1	99.3	99.4	99.7	99.3	98.6
M2	98.7	99.2	99.3	99.8	98.8
M3	98.6	99.7	99.6	99.2	99.2
M4	98.9	98.9	99.2	98.8	99.4
M5	98.7	98.6	99.8	98.5	98.7
M6	99.4	99.5	98.5	98.9	99.6
M7	98.7	99.3	98.7	99.3	98.9
M8	98.8	98.7	98.9	99.7	99.3
M9	99.5	98.7	99.3	99.3	99.6
M10	98.7	99.3	98.8	98.6	99.2
Mean	98.93	99.13	99.18	99.14	99.13

Table 8 is the experimental analysis of MEMD-CNN for meningioma detection system on BRAINWEB dataset. The proposed MEMD-CNN method for meningioma classification approach attains 99.36% of SET, 99.41% of SPT, 99.62% of MSA, 99.53% of PR and 99.57% of FS.

Table 8

Experimental analysis of MEMD-CNN for meningioma detection system

Meningioma	Experimental metrics in %
image sequences	SET	SPT	MSA	PR	FS
M1	99.56	99.4	99.7	99.3	99.5
M2	99.3	99.2	99.7	99.8	99.8
M3	99.2	99.7	99.8	99.2	99.2
M4	99.3	99.4	99.2	99.5	99.9
M5	99.1	99.5	99.8	99.8	99.6
M6	99.4	99.5	99.7	99.7	99.7
M7	99.3	99.3	99.8	99.3	98.9
M8	99.7	99.3	99.5	99.7	99.7
M9	99.5	99.5	99.3	99.3	99.6
M10	99.3	99.3	99.7	99.7	99.8
Mean	99.36	99.41	99.62	99.53	99.57

Table 9 is the experimental analysis of meningioma detection system using EMD and MEMD approaches on both BRAINWEB and Nanfang university datasets. The proposed meningioma detection system using conventional EMD transformation approach achieves 99.14% of SET, 94.89% of SPT, 95.17% of MSA, 94.29% of PR and 93.85% of FS on Nanfang university dataset. Also, the proposed meningioma detection system using MEMD transformation approach achieves 98.93% of SET, 99.13% of SPT, 99.18% of MSA, 99.14% of PR and 99.13% of FS on Nanfang university dataset.

Table 9

Experimental analysis of meningioma detection system using EMD and MEMD approaches on BRAINWEB and Nanfang university datasets

Experimental metrics in %	Nanfang university dataset		BRAINWEB dataset
	Meningioma detection system with conventional EMD approach	Meningioma detection system with proposed MEMD approach	Meningioma detection system with conventional EMD approach	Meningioma detection system with proposed MEMD approach
SET	94.14	98.93	93.98	99.36
SPT	94.89	99.13	93.28	99.41
MSA	95.17	99.18	94.75	99.62
PR	94.29	99.14	94.28	99.53
FS	93.85	99.13	94.39	99.57

The proposed meningioma detection system using conventional EMD transformation approach achieves 93.98% of SET, 93.28% of SPT, 94.75% of MSA, 94.28% of PR and 94.39% of FS on BRAINWEB dataset. Also, the proposed meningioma detection system using MEMD transformation approach achieves 99.36% of SET, 99.41% of SPT, 99.62% of MSA, 99.53% of PR and 99.57% of FS on BRAINWEB dataset.

Table 10 is the comparisons of proposed MEMD-CNN method for meningioma detection system with conventional methods Irmak et al. (2021), Balakumaresan Ragupathy et al. (2020) and Selvapandian et al. (2018). From Table 10, the meningioma detection system using MEMD-CNN method significantly obtains high performance metrics than the conventional meningioma detection approaches.

Table 10

Comparisons of proposed MEMD-CNN method for meningioma detection system with conventional methods

Methods	SET	SPT	MSA	PR	FS
(%)	(%)	(%)	(%)	(%)
Proposed MEMD-CNN approach	98.93	99.13	99.18	99.14	99.13
Irmak et al. (2021)- Deep Convolutional Neural Network	97.10	97.82	96.86	97.04	97.18
Balakumaresan Ragupathy et al. (2020)- U-Net CNN classification	96.26	96.64	97.18	96.26	97.38
Selvapandian et al. (2018)- Gradient boosting classifier	97.28	96.96	97.79	97.38	97.49

Table 11 is the comparisons of proposed MEMD-CNN method for meningioma detection system with conventional methods Irmak et al. (2021), Balakumaresan Ragupathy et al. (2020) and Selvapandian et al. (2018). From Table 11, the meningioma detection system using MEMD-CNN method significantly obtains high performance metrics than the conventional meningioma detection approaches.

Table 11

Comparisons of proposed MEMD-CNN method for meningioma detection system with conventional methods on BRAINWEB dataset

Methods	SET	SPT	MSA	PR	FS
	(%)	(%)	(%)	(%)	(%)
Proposed MEMD-CNN approach	99.36	99.41	99.62	99.53	99.57
Irmak et al. (2021)- Deep Convolutional Neural Network	93.91	95.29	95.37	94.85	95.38
Balakumaresan Ragupathy et al. (2020)- U-Net CNN classification	97.17	98.37	97.28	97.74	97.86
Selvapandian et al. (2018)- Gradient boosting classifier	98.01	98.26	98.17	97.38	97.38

5 Conclusions

The meningioma images are detected from non-meningioma brain images using MEMD-CNN classification method. The meningioma detection system using conventional EMD transform obtains 92.9% of MCR and 97.6% of NMCR, whereas the meningioma detection system using MEMD transform obtains 99.4% of MCR and 99.3% of NMCR on the meningioma and non-meningioma images in constructed dataset. The proposed meningioma detection system using conventional EMD transformation approach achieves 99.14% of SET, 94.89% of SPT, 95.17% of MSA, 94.29% of PR and 93.85% of FS on Nanfang university dataset. Also, the proposed meningioma detection system using MEMD transformation approach achieves 98.93% of SET, 99.13% of SPT, 99.18% of MSA, 99.14% of PR and 99.13% of FS on Nanfang university dataset. The proposed meningioma detection system using conventional EMD transformation approach achieves 93.98% of SET, 93.28% of SPT, 94.75% of MSA, 94.28% of PR and 94.39% of FS on BRAINWEB dataset. Also, the proposed meningioma detection system using MEMD transformation approach achieves 99.36% of SET, 99.41% of SPT, 99.62% of MSA, 99.53% of PR and 99.57% of FS on BRAINWEB dataset.

References

Cinar

and Yildirim

, Detection of tumors on brain MRI images using the hybrid convolutional neural network architecture, Med Hypotheses 139 (2020), 109684.

Deepak

and Ameer

, Brain tumor classification using deep CNN features via transfer learning, Comput Biol Med 111 (2019), 103345.

Kabir Anaraki

, Ayati

and Kazemi

, Magnetic Resonance imaging-based brain tumor grades classification and grading via convolutional neural networks and genetic algorithms, Biocybern Biomed Eng 39(1) (2019), 63–74.

Mehrotra

, Ansari

M.A.

, Agrawal

and Anand

R.S.

, A Transfer learning approach for AI-based classification of brain tumors, Mach Learn Appl 2(9) (2020), 1–12.

Muhammad

, Khan

, Ser

J.D.

and Albuquerque

V.H.C.

, Deep learning for multigrade brain tumor classification in smart healthcare systems: A prospective survey, IEEE Trans Neural Netw Learn Syst 32(2) (2021), 507–522.

Rehman

, Naz

, Razzak

M.I.

, Akram

and Imran

, A Deep Learning-based framework for automatic brain tumors classification using transfer learning, Circuits, Syst, Signal Process 39(2) (2020), 757–775.

Shirazi

A.Z.

, Fornaciari

, McDonnell

M.D.

, Yaghoobi

, Cevallos

, Tello-Oquendo

, Inca

and Gomez

G.A.

, The application of deep convolutional neural networks to brain cancer images: A survey, J Personal Med 10(4) (2020), 1–27.

Tiwari

, Srivastava

and Pant

, Brain tumor segmentation and classification from magnetic resonance images: Review of selected methods from to, Pattern Recogn Lett 131 (2020), 244–260.

Prilianti

, Brotosudarmo

, Anam

, Suryanto

Performance comparison of the convolutional neural network optimizer for photosynthetic pigments prediction on plant digital image, AIP Conference Proceedings (2019).

10.

Ragupathy

and Karunakaran

, A fuzzy logic-based meningioma tumor detection in magnetic resonance brain images using CANFIS and U-Net CNN classification, International Journal of Imaging Systems and Technology 31(1) (2020).

11.

Selvapandian

and Manivannan

, Performance analysis of meningioma brain tumor classifications based on gradient boosting classifier, International Journal of Imaging Systems and Technology 28(4) (2018).

12.

Irmak

, Multi-classification of brain tumor MRI images using deep convolutional neural network with fully optimized framework, Iran J Sci Technol Trans Electr Eng 45 (2021), 1015–1036.

13.

Jasmine Hephzipaha

and Thirumurugan

, Performance analysis of meningioma brain tumor detection system using feature learning optimization and ANFIS classification method, IETE Journal of Research 1(3) (2020), 1–12.

14.

Song

, Shan

, Baod

, Liud

, Zhao

and Chen

, Automatic brain tumour diagnostic method based on a back propagation neural network and an extended set-membership filter, Computer Methods and Programs in Biomedicine 208 (2021), 106188.

15.

BrainWeb dataset, Available at: https://brainweb.bic.mni.mcgill.ca/

16.

https://doi.org/10.6084/m9.figshare.1512427.v5.