Detection and Segmentation of Glioma Tumors Using an Improved Visual Geometry Group (IVGG) Deep Learning Structure

Abstract

Glioma brain tumors have similar textural patterns to other tumors, making their detection and segmentation a challenging process. The approach of the Modified Tumor Detection System (MTDS) is presented in this study to identify and categorize brain images of gliomas from images of healthy brains. The Spatial Gabor Transform (SGT), feature calculations, and deep learning structure comprise the training work flow of the suggested MTDS technique. The features are computed from the glioma brain image dataset images and the normal brain image dataset images and these features are fed into the classification architecture. In this paper, the proposed IVGG architecture is derived from the existing Visual Geometry Group (VGG) architecture to improve the detection rate of the proposed system and to decrease the computational time complexity. The testing work flow of the proposed system is also consist of SGT, feature computation and the IVGG architecture to produce the classification result of the source brain images into either normal or glioma. Furthermore, the Morphological Segmentation technique has been used to find the tumor locations in this glioma image. Two separate brain imaging datasets have been used in this study to evaluate and validate the suggested MTDS's performance efficiency. BRATS Imaging 2020 (BI20) and Kaggle Brain Imaging (KBI) are the datasets. Analysis of the performance efficiency has been done in relation to the Jaccard index, recall, precision, and detection rate.

Keywords:
MTDS; glioma; brain tumors; deep learning; IVGG

HIGHLIGHTS

The proposed IVGG architecture is proposed for brain tumor detection.

The proposed method on the brain images from BRATS Imaging 2020 and Kaggle Brain Imaging.

The performance of decrease the computational time, MTDS's performance efficiency is analyzed.

INTRODUCTION

Every day, the human body's nervous system plays a crucial role in human activity. The nervous system in human body plays an important role in human activity every day. The brain, a vital organ in the human body responsible for all physical and mental activities, plays a crucial role within the nervous system. Any aberration in the brain reverberates throughout the entire nervous system, significantly impacting bodily functions. Numerous abnormalities manifest in the human brain, among which tumors hold paramount importance due to their life-threatening nature, gradually compromising brain functions based on the body's immune response. Uncontrolled tissue growth in the brain leads to tumor formation, a condition contributing to around 250,000 annual fatalities, as reported by the World Health Organization (WHO) and the US Brain Tumor Academy. These tumors are classified by presence, size, shape, and intensity, with categories including Glioma, Glioblastoma, and pituitary tumors, according to WHO standards. Gliomas, for instance, arise from glial cells, which provide support and protection for neurons in the brain. These tumors can be further classified into subtypes such as astrocytomas, oligodendrogliomas, and ependymomas, each with varying degrees of malignancy. Glioblastomas are the most aggressive form of gliomas, characterized by rapid growth and a poor prognosis. Pituitary tumors, on the other hand, develop in the pituitary gland, a small but crucial gland located at the base of the brain responsible for hormone production and regulation. While many pituitary tumors are benign, they can still cause significant health issues by disrupting hormone balance and exerting pressure on surrounding brain structures. The early detection and accurate classification of brain tumors are vital for effective treatment and improving patient outcomes. Advances in medical imaging technologies such as Magnetic resonance imaging (MRI), Computed tomography (CT) scans, and Positron emission tomography (PET) scans have greatly enhanced the ability to diagnose and monitor these tumors. Moreover, the integration of machine learning and deep learning techniques in medical diagnostics holds promise for even more precise and efficient identification of brain tumors, facilitating timely and appropriate interventions. Understanding the complexities of brain tumors and their impact on the nervous system is essential for developing better diagnostic tools and therapeutic strategies. Continued research and technological advancements are crucial in the fight against this formidable health challenge, aiming to reduce mortality rates and improve the quality of life for those affected by brain tumors. Gliomas originate from glial cells, which are essential for supporting and protecting neurons in the brain. These tumors are categorized into various subtypes, including astrocytomas, oligodendrogliomas, and ependymomas, each presenting different levels of malignancy. Glioblastomas represent the most severe type of glioma, known for their rapid progression and poor prognosis. Conversely, pituitary tumors form in the pituitary gland, a small yet vital gland at the base of the brain that regulates hormone production. Although many pituitary tumors are benign, they can still lead to serious health issues by disrupting hormonal balance and exerting pressure on nearby brain structures. Timely detection and precise classification of brain tumors are crucial for effective treatment and improving patient outcomes. Medical imaging technologies such as MRI, CT scans, and PET scans have significantly advanced the diagnostic and monitoring capabilities for these tumors. Additionally, incorporating machine learning and deep learning techniques into medical diagnostics offers the potential for even more accurate and efficient identification of brain tumors, enabling prompt and suitable interventions. Comprehending the intricacies of brain tumors and their effects on the nervous system is vital for developing enhanced diagnostic tools and therapeutic strategies. Ongoing research and technological innovations are essential in addressing this significant health challenge, with the goal of reducing mortality rates and enhancing the quality of life for individuals affected by brain tumors.

Glioma images, such as the one in Figure 1(a) represents the MRI image of healthy brain, while Figure 1(b) represents the MRI image of tumor affected brain.

Figure 1(a)
Healthy brain MRI

Figure 1(b)
Brain tumor MRI

Proper diagnosis methods can extend the lifespan of brain tumor patients by facilitating appropriate medication. It is vital for skilled radiologists or doctors with specialized training in this area to identify problems early. X-rays, CT scans, MRIs, PET scans, and magnetic resonance imaging (MRI) are among the scanning techniques that are used to diagnose and evaluate brain cancers [¹-⁴]. Radiation exposure, pixel intensity, and imaging quality vary across these methods. X-rays, despite using low radiation, have limited accuracy and are unsuitable for further processing. Conversely, PET, employing high radiation, isn't safe for patients. This research opts for MRI scanning due to its high pixel accuracy, aiding in locating tumor regions within the brain images [⁴]. However, manual detection is time-consuming, prompting the necessity to automate the process using computer-aided simulation methods, employing machine and deep learning algorithms. Here, tumor areas in brain MRI scans are recognized and highlighted using an improved deep learning algorithm. The outline of the paper is as follows: Section 2 examines current glioma detection techniques and the outcomes of their experiments, Section 3 introduces a glioma detection system, Section 4 details experimental outcomes, and Section 5 provides a comprehensive summary of the research, including limitations and future scopes.

LITERATURE SURVEY

Kokila B and coauthors (2021) developed a model to use MRI to identify brain cancers. It entails detecting the tumor, determining its nature and grade, and determining its exact location [⁵]. Our method grouped brain MRI data for many classification tasks using a single model, as opposed to utilizing different models for each task. Since Convolutional Neural Network (CNN) has the ability to classify and detect tumors, themultitask classification process depends on these capabilities. Brain tumor locations can also be identified with 92 percent accuracy using an Advanced Convolutional Neural Network (ACNN) based algorithm.

Amin J and coauthors (2020) The combined images are sent into CNN so that it can automatically identify features and classify brain tumors [⁶]. On fused images, the approach generates the best results: 0.97 ACC on the datasets for the BRATS 2012 Image, 0.98 ACC on the BRATS 2013 Challenge, 0.96 ACC on the BRATS 2013 Leader board, 1.00 ACC on the BRATS 2015 Challenge, and 0.97 ACC on the BRATS 2018 Challenge.

Prasad G and coauthors(2023) Machine learning classifiers employ the GLCM characteristics that were taken out of the machined surface images as inputs [⁷]. The DRT's sensitive characteristics are chosen using a threshold criterion function (TCF). With a classification accuracy of up to 95.3%, the Random Forest (RAF) model outperformed all the other classifiers. The analysis's findings also demonstrate how well the most sensitive features are identified by the suggested dimensionality reduction approach using TCF. A maximum decrease of 62% in dimensionality is attained. Compared to the methods described in the prior study, the suggested methodology demonstrated a 7.2% increase in classification accuracy.

Hephzipah JJ and coauthors (2020) After applying the Adaptive Neuro Fuzzy Inference System (ANFIS) classification methodology to these optimized features, the inappropriate meningioma brain image is diagnosed [⁸]. The tumor regions are then segmented using the morphological segmentation method. 98.1% sensitivity, 99.75 specificity, 99.6% accuracy, 98.55 precision, 97.95 F1-Score, and 98.1% relevance factor have been achieved by the recommended meningioma tumor detection approach.

Fayaz M and coauthors (2021) The images used in the suggested model have undergone 3-level Harr wavelet decomposition to eliminate fine details and condense their size [⁹]. Convolutional neural networks are then utilized to categorize brain MR images into both healthy and unhealthy regions. Another popular classification technique that has been applied extensively in many fields is the convolutional neural network. This study classified brain MRI images using a convolutional neural network.

The segmented tumor regions have been further diagnosed in their work based on their placements. Sasikanth S and coauthors (2017) developed an automated system for the detection and segmentation of glioma tumors using a graph cut-based ANFIS classifier technique [¹⁰]. They have considered features such local binary patterns, local ternary patterns, Laws texture, and GLCM that were obtained from the enhanced image via oriental analysis and subsequently classified using ANFIS.

Madhukumar S and coauthors (2015) this system separated the training dataset automatically using a watershed methodology, avoiding the need for initial human segmentation [¹¹]. This increased segmentation accuracy from 82 to 89%, and it might be useful for CNN development in the future. There wasn't much information here about the specifics of the convolution layer or picture filters. This is also the reason that feature extraction uses a wide range of techniques and algorithms. This study proposes a unique IVGG structure for the purpose of identifying and segmenting the tumor regions in MRI images of the brain associated with glioma.

Mohamed and coauthors (2020) demonstrates the potential of using deep learning in MRI images to provide a non-invasive tool for simultaneous and automated tumor segmentation, detection, and grading of LGG for clinical applications. [¹²]

Atika Akter and coauthors (2024) classification model achieved the highest accuracy of 98.7% in a merged dataset and 98.8% with the segmentation approach, with the highest classification accuracy reaching 97.7% among the four individual datasets. [¹³]

Kurc and coauthors (2020) developed three classification methods to categorize adult diffuse glioma cases into oligodendroglioma and astrocytoma classes using radiographic and histologic image data. These methods achieved accuracy values of 0.75, 0.80, and 0.90, measured as the ratio of the number of correct classifications to the number of total cases, with the challenge dataset. [¹⁴]

Aurna NF and coauthors (2022) proposed two stage ensemble model is analyzed using several performance metrics and three different experiments. Through the prominent performance, the proposed model is able to outperform other existing models attaining an average accuracy of 99.13% by optimization of the developed algorithms. Here, the individual accuracy for Dataset 1, Dataset 2, Dataset 3, and Merged Dataset is 99.67%, 98.16%, 99.76%, and 98.96% respectively. Finally a User Interface (UI) is created using the proposed model for real time validation. [¹⁵]

Díaz-Pernas FJ and coauthors (2021) proposed neural model can analyze MRI images containing three types of tumors: meningioma, glioma, and pituitary tumor, over sagittal, coronal, and axial views and does not need preprocessing of input images to remove skull or vertebral column parts in advance. The performance of our method on a publicly available MRI image dataset of 3064 slices from 233 patients is compared with previously classical machine learning and deep learning published methods. [¹⁶]

Aurna NF and coauthors (2021) proposed feature level ensemble of 3 CNN models increases the model robustness and efficacy to a great extent. Further, Analysis of Principal Component (PCA) is done for feature or dimensionality reduction which also improves the performance of the model considering execution time and accuracy which shows a prominent performance that outperforms other existing models along with the pre-trained models obtaining an average validation accuracy of 98.37%.[¹⁷]

Ahmad and coauthors (2019) proposed algorithm obtained satisfactory results on healthy and pathological liver CT images. Our algorithm achieved 94.80% Dice similarity coefficient on mixed (healthy and pathological) images while 91.83% on pathological liver images, which is better than those of the state-of-the-art methods. [¹⁸]

The SGT transformation process is integrated with the deep learning structure to improve the detection rate of the proposed MTDS approach.

Most of the conventional glioma tumor detection methods applied their methods on either low or high intensity glioma brain image dataset. The methodologies stated in this research work are evaluated on both low and high intensity glioma brain image dataset.

PROPOSED METHODOLOGIES

In order to identify and categorize brain images with gliomas from those with healthy brains, a fully automated soft computing methodology based MTDS approach is presented in this article. The suggested MTDS solution uses SGT, feature calculations, and a deep learning framework as its training workflow. The normal brain image dataset and the glioma brain image dataset are used to generate the features, which are then supplied into the classification framework. The suggested IVGG architecture in this work, which attempts to lower computational time complexity and raise the suggested system's recognition rate, is based on the existing VGG design. SGT, feature computation, and IVGG architecture are also included in the suggested system's testing procedure to produce a classification result for the source brain images into normal and glioma categories. Additionally, the tumor areas in this glioma image have been identified using the application of the Morphological Segmentation approach.

Figure 2(a) illustrates the recommended MTDS strategy, and Figure 2(b) illustrates the tumor segmentation method. Data augmentation techniques are applied in the training workflow on the brain images that are a part of the training dataset in order to raise the count values of the brain images and get a higher detection rate. In this work, brain images that are limited to the training dataset are enhanced using shift pixel left and shift pixel right data augmentation approaches. There is no data augmentation module in the testing protocol of the proposed MTDS solution.

Figure 2(a)
Proposed MTDS approach

Figure 2(b)
Tumor segmentation approach

Spatial Gabor Transform

The spatial to frequency transformation process is done by multi resolution transformation models. These transforms are categorized into Gabor, Contourlet and Curvelet. Among these transformational models, Contourlet and Curvelet transform exhibits loss of pixel components during its transformation stages, which reflect in the reduction of the detection rate of the glioma classification system. Hence, this research work uses SGT (for performing the spatial to frequency transformation in the source brain image. This SGT works based on the spatial kernel and this two dimensional kernel is multiplied with the source brain MRI image. The kernel of this SGT is given in the below equation (1).

$Kernel\left(SGT\right)=\exp \left(\frac{-1}{2}{\left(\frac{p^{|}}{s*p^{|}}\right)}^2+{\left(\frac{q^{|}}{s*q^{|}}\right)}^2\right)*\mathrm{Cos}2\pi f{p}^{|}$ (1)

Whereas, $p^{|}$ and $q^{|}$ are the transformational pixel coordinates with scaling parameter s and frequency factor ‘f’.

These transformational pixel coordinates are given in the below equations with respect to its pixel orientation. The orientation of each pixel is referred by the variable ‘theta’ and the value of this ‘theta’ is varied between -90 degrees to +90 degrees. The transformational pixel coordinates' functional properties are explained by the following equations.

$p^{|}=p*\cos \left(theta\right)+q\sin (theta)$

$q^{|}=q*\cos \left(theta\right)-p\sin (theta)$

In this paper, the function of this SGT is determined by the scaling parameter and pixel orientation theta values. The value of scaling parameter is set to unique and hence the SGT stated in this work is entirely depending on the value of the theta. In this research work, the value of theta decides the characteristics of the SGT in the glioma brain image classification process. Hence, totally 180 number of SGT values are generated for the entire theta values which is processed through the transformational pixel coordinates. These 180 SGT values are now convolved with the source brain image to produce 180 number of SGT images. The unique SGT image is generated by combining all these 180 SGT images by selecting the highest pixel intensity in each SGT image. From this unique SGT image, features are computed and classified by the proposed classification structure stated in this paper.

Feature computation process

The outcomes of calculating or removing the most crucial data from an image are its features. The process of extracting features from an image is referred to as the "feature computation process". Using features taken from various brain imaging regions of interest, radiologists can differentiate between normal and abnormal images. In this research work, Grey Level CooccurrenceMatrix (GLCM) (Bhavani R and coauthors, 2023) and Local Ternary Pattern (LTP) features (Shamna N V and coauthors 2023) are computed from the SGT image and these features are fed into the classifier to perform the classification process [¹⁹-²⁰]. The feature computation process is explained in the following sections.

GLCM

The differences between the abnormal image and the normal image are highlighted by the variances in pixels. The matrix known as GLCM is constructed with respect to different phases of pixel orientations and is used to compute specific grey level information from the image. This matrix is created in this paper taking into account the 45-degree pixel orientation phase. This matrix is now used to compute the following grey level features as described in the below equations 2, 3, 4, 5, 6, 7, and 8.

$Energy={\displaystyle \sum }_{i=1}^P{\displaystyle \sum }_{j=1}^Q{M}_{ij}{}^2$ (2)

Whereas, $M_{ij}$ is the GLCM with respect to row value $i$ and column value $j$. P and Q are rows and columns count.

$Entropy={\displaystyle \sum }_{i=1}^P{\displaystyle \sum }_{j=1}^Q{M}_{ij}*\log (M_{ij})$ (3)

$Homogeneity={\displaystyle \sum }_{i=1}^P{\displaystyle \sum }_{j=1}^Q\frac{M_{ij}}{1+|i-j|}$ (4)

$AutoCorrelation={\displaystyle \sum }_{i=1}^P{\displaystyle \sum }_{j=1}^{Q}i*j*M_{ij}$ (5)

$Dissimilarity={\displaystyle \sum }_{i=1}^P{\displaystyle \sum }_{j=1}^Q{M}_{ij}*|i-j|$ (6)

$ClusterShade={\displaystyle \sum }_{i=1}^P{\displaystyle \sum }_{j=1}^Q{M}_{ij}*(i+j-{\mu }_i-{\mu }_j)$ (7)

$Cluster\mathrm{Pr}o\min ence={\displaystyle \sum }_{i=1}^P{\displaystyle \sum }_{j=1}^Q{M}_{ij}*(i+j-{\mu }_i-{\mu }_j)$ (8)

LTP

This binary set of features are computed for each pixel in the SGT image and it is derivd from Local Binary Pattern (LBP) feature. The two features are generated for each pixel in LTP computation process. The following equation9 is used to compute the LTP feature.

$LTP\left(x_c\right)=\{\begin{array}{c}+1;x_f>x_c+k\\ 0;|x_f-x_c|<k\\ -1;x_f\le x_c-k\end{array}$ (9)

Whereas, $x_c$ is the pixel in the SGT image where the feature value is to be computed and $x_f$ is the surrounding pixels in 3*3 mask which is placed over the SGT image to compute LTP feaure values. $\text{'}k\text{'}$ is the kappa index, which has the random value between 0 and 255. In this research work, the value of kappa index is set to low value in order to reduce the overfitting errors during the LTP feature computing process.

Classifications

Classification receives the computed features from both normal brain image and the abnormal brain image and produces the training vectors during the training phase of this proposed classifier. Many experts in this medical imaging field used deep learning algorithms for the classification of brain tumors from the past decade. The existing deep learning architectures are LeNet, AlexNet and VGG. Among these existing deep learning structures, the VGG is used to obtain the higher detection rate of the glioma classification system (Parameswari and coauthors 2024) [²¹]. Therefore, VGG and its improved version structures are used in this paper to identify or differentiate the glioma brain images from the healthy brain images. In existing VGG structure for glioma image classification process (Anita JN and coauthors 2022) [²²], the structure consist of thirteen numbers of Convolutional layers and five numbers of pooling layers with two numbers of dense layers, as illustrated in Figure 3 (a). This existing structure is split into five modules where the modules 1 and 2 consists of two Convolutional layers respectively and modules 3, 4 and 5 consists of three Convolutional layers respectively. The two Convolutional layers in module 1 is designed with 32 filters, the two Convolutional layers in module 2 is designed with 64 filters, three Convolutional layers in module 3 is designed with 128 filters, three Convolutional layers in module 4 is designed with 256 filters and three Convolutional layers in module 5 is designed with 512 filters. The module 5 output is passed to two consecutive dense layers Dense1 and and Dense3. The Dense 3 output is either Glioma or healthy. The output result is entirely depending on the number of features generated internally. The generation of internal features within this structure is depending on the number of Convolutional layers and the number of filters in each Convolutional layers. In this existing structure, each Convolutional layer is designed with less number of filters and hence they produced less number of internal features. Moreover, the negativity responses from each Convolutional layer output are not removed in this existing structure which further reduces the detection rate of the entire proposed system. Hence, this research work proposes a highly efficient deep learning structure which is the improved version of the conventional VGG structure.

Proposed Algorithm for Glioma Classification Using IVGG

Step 1: Input Preparation

Input: Features extracted from normal and abnormal brain images.

Step 2: IVGG Architecture

Module 1:

Two Convolutional layers with 64 filters each.

ReLU activation to remove negative values.

Pooling to reduce feature map size.

Module 2:

Two Convolutional layers with 128 filters each.

ReLU activation and pooling similarly applied.

Parallel Convolutional Layers:

Module 1: Three Convolutional layers with 256 filters each.

Module 2: Three Convolutional layers with 512 filters each.

ReLU activation after each layer to enhance features.

Additional pooling to further condense feature maps.

Step 3: Feature Map (FM) Construction

Combine features from Module 1 and Module 2 to form a unified Feature Map (FM).

Step 4: Classification

Dense Layers:

Pass FM through Dense layers (Dense1, Dense2).

Output from Dense2 determines if the image is Glioma or Healthy.

Step 5: Output

Output: Final classification result (Glioma or Healthy).

The Figure 3 (b) represents Improved VGG (IVGG) structure in the proposed system consists of two parallel modules module 1 and module 2. The parallel module 1 consists of two Convolutional layers and each having 64 filters. The parallel module 2 consists of two Convolutional layers and each having 128 filters. The negativity responses from the output of Convolutional layers C2 and C4 in Fig. 3(b), are removed by passing the output through Rectification Linear Unit (ReLU). The negativity removed responses are now size reduced by pooling layers (P). The final output response from the module 1 is passed to the three Convolutional layers C5, C6 and C7 in parallel manner and each having 256 numbers of filters. The final output response from the module 1 is passed to the three Convolutional layers C8, C9 and C10 in parallel manner and each having 512 numbers of filters. Then, the negativity of output responses from all these Convolutional layers are removed by passing them through the ReLU layer. The output of the ReLU layer is size reduced by pooling layers (P). The intrinsic features from module 1 are generated by accumulating the final responses from pooling layers 3, 4 and 5. The intrinsic features from module 2 are generated by accumulating the final responses from pooling layers 6, 7 and 8. The Feature Map (FM) is constructed by accumulating both intrinsic features.

Figure 3(a)
Conventional VGG-16 structure

Figure 3(b)
Improved VGG (IVGG) structure

Figure 3(c)
Block diagram of Improved VGG (IVGG)

The Figure 3 (c) represents The Improved VGG architecture incorporates several enhancements to the original VGG model, aiming to improve its performance and feature extraction capabilities. Here's an in-depth explanation of each component:

SE Blocks (Squeeze-and-Excitation)

SE Blocks are integrated into Block1 to Block6 of the VGG architecture. These blocks are designed to adaptively recalibrate channel-wise feature responses, enhancing the network's ability to focus on informative features and suppress irrelevant ones. The SE Block consists of three main components:

Squeeze (Global Average Pooling):

Initially, the input feature map (HxWxC) undergoes global average pooling across the spatial dimensions (H and W). This step reduces the spatial dimensions to 1x1 while retaining channel-wise information. The result is a set of channel-wise descriptors representing the global statistics of each feature map.

Excitation (Fully Connected Network - FCN):

Following global average pooling, a small fully connected network (or 1x1 convolutions) is applied. This network consists of two layers:

The first layer reduces the number of channels using a reduction ratio (typically smaller than the original number of channels), followed by a non-linear activation function (e.g., ReLU).

The second layer restores the original number of channels and applies a sigmoid activation function. This produces a set of scaling factors (weights) for each channel, representing the importance of each channel's response.

Recalibrate (Multiply with SE Weights):

The recalibration step involves applying the learned scaling factors (weights) to the original feature map. Each channel is multiplied by its corresponding weight, enhancing the informative channels and suppressing less relevant ones.

The output of the SE Block is a recalibrated feature map (HxWxC), which retains the spatial dimensions and channel information from the input but with enhanced channel-wise responses.

Self-Attention Layer

After the SE Block, a Self-Attention Layer is introduced. This layer plays a crucial role in capturing long-range dependencies within the feature maps. It allows the network to selectively focus on important features across different spatial locations, enhancing feature extraction capabilities. By attending to relevant spatial locations based on learned attention weights, the Self-Attention Layer improves the network's ability to understand complex relationships within the input data.

Mix-up Augmentation

Following the Self-Attention Layer, a Mix-up Layer is included to perform mix-up data augmentation. Mix-up augmentation generates new training samples by blending pairs of images and their corresponding labels. This technique helps regularize the model and improve its generalization ability by exposing it to diverse combinations of data points. By blending images and labels, the Mix-up Layer encourages the model to learn more robust features and reduces overfitting.

Integration with CNN

The Improved VGG architecture integrates these components seamlessly into the Convolutional Neural Network (CNN) framework. Typically, SE Blocks are placed after convolutional layers in Blocks 1 to 6 of the VGG model. This integration enhances the network's capability to capture significant channel-wise relationships and long-range dependencies within the input data. As a result, the Improved VGG model demonstrates improved performance across various computer vision tasks, including image classification, object detection, and semantic segmentation.

By incorporating SE Blocks for adaptive feature recalibration, Self-Attention Layers for capturing dependencies, and Mix-up Augmentation for data diversity, the Improved VGG architecture advances the state-of-the-art in deep learning by enhancing both feature extraction and model generalization capabilities. These enhancements collectively contribute to better accuracy and robustness in handling complex visual tasks.

The output response from this FM is now passed to two consecutive dense layers Dense1 and Dense2 The Dense 2 output is either Glioma or healthy. The main difference between the existing and proposed structure is that the proposed structure is designed with more number of filters in each Convolutional layer which produces high number of internal features. This improves the detection rate of the entire proposed system.

Figure. 4(a) is the IVGG classification results of healthy brain image and Figure. 4(b) is the IVGG classification results of the glioma brain image respectively.

Figure 4(a)
IVGG results of Healthy images

Figure 4(b)
IVGG results of Glioma images

Segmentation

The tumor region point of pixels in the glioma brain image is identified or located using segmentation process or algorithm. In this research work, Morphological Algorithm (MA) (Arash Rabbani and coauthors 2023) is used to perform the segmentation process. Morphological algorithms for image segmentation leverage mathematical morphology, which is a theoretical approach for examining shapes and structures within images. These algorithms perform operations that manipulate images based on their spatial configuration, often working with binary or grayscale images. The core morphological operations are dilation, erosion, opening, and closing, each of which modifies the image's geometric structure in a distinct manner.This MA approach for locating the tumor pixels has Morphological Opening (MO) and Morphological Closing (MC) [²³]. The MO and MC are otherwise called as Morphological Dilation (MD) and Morphological Erosion (ME). During MD process, the pixels in the glioma image has been enhanced at 3mm radius using disk shaped structuring function. During ME process, the pixels in the glioma image has been shrink at 2mm radius using disk shaped structuring function. Now, the subtraction is performed between the MD and ME images in order to segment the tumor region point of pixels in this paper. The source brain MRI picture is displayed in Figure 5(a), and the tumor-segmented brain MRI image, where the tumor regions are found by the segmentation algorithm, is displayed in Figure 5(b).

Figure 5(a)
Brain MRI image

Figure 5(b)
Tumor segmented brain MRI image

RESULTS AND DISCUSSIONS

In this study, the effectiveness of the proposed MTDS (Modified Tumor Detection System) was assessed using two independent brain imaging datasets, KBI and BI20. These datasets contain open-access brain MRI images available for global research use but restricted for commercial use without proper licensing.

The KBI dataset comprises 3064 T1-enhanced brain MRI images from 233 patients aged between 10 and 80, categorized into glioma, Glioma, and Pituitary tumor types. Among these, there are 708 images for glioma, 1426 for Glioma, 930 for pituitary tumors, and 1250 healthy brain images, all sized at 512512 pixels. The BI20 dataset contains 2400 healthy brain images, 1099 glioma type images, and 785 Glioma type images, all sized at 10241024 pixels. All images were annotated by three independent subject experts in radiology.

The datasets were divided into Training Modules (TrM) and Testing Modules (TeM) using a 70:30 ratio to validate the proposed MTDS. For instance, in KBI, 70% of glioma images were allocated to TrM (Training Module), and the remaining 30% to TeM (Testing Module). Similar ratios were used for healthy brain images.

The proposed MTDS achieved impressive detection ratios (DR) of 99% for glioma and 98.9% for healthy images in the KBI dataset, averaging to 98.95% DR. In the BI20 dataset, it achieved a DR of 98.7% for glioma and 99.7% for healthy images, averaging a DR of 99.2%.

Comparing the modified and existing VGG architectures in Table 1, the existing VGG architecture achieved 97.81% DR on KBI dataset brain images and 96.07% on BI20 dataset images. Meanwhile, the IVGG (Improved VGG) in the proposed system achieved notably higher DRs of 98.95% on KBI and 99.2% on BI20 dataset brain images, indicating the superior performance of the proposed architecture in both datasets. KBI and BI20 datasets in the classification module of the proposed system.

Table 1
DR comparisons between modified and existing VGG architecture

Datasets	DR in %
	Proposed IVGG	Existing VGGarchitecture
KBI	98.95	97.81
BI20	99.2	96.07

Table 2 is the impact analysis of data augmentation in the proposed MTDS by implementing data augmentation methods in TrM of the proposed classification system in this paper. The proposed MTDS with data augmentation in the modified VGG architecture attains 98.95% of DR in KBI dataset and also attains 99.2% of DR in BI20 dataset, whereas the proposed MTDS without data augmentation in the modified VGG architecture attains 96.98% of DR in KBI dataset and also attains 97.75% of DR in BI20 dataset.

Table 2
Impact analysis of data augmentation in the proposed MTDS

Datasets	DR in %
	IVGG with data augmentation in TrM	IVGG without data augmentation in TrM
KBI	98.95	96.98
BI20	99.2	97.75

With the analysis of DR for the proposed MTDS approach, the following equations are also used to describe the functional effectiveness of the glioma classification system.

When the model accurately predicts the positive class-the existence of a brain tumor-these are known as True Positives (TP).

When a brain tumor is not present despite the model predicting a positive class, this is known as a False Positive (FP).

$\mathrm{Pr}ecision=\frac{TruePositives}{TruePositives+FalsePositives}$ (10)

When a brain tumor is present despite the model's prediction of a negative class, these situations are known as False Negatives (FN).

$\mathrm{Re}call=\frac{TruePositives}{TruePositives+FalseNegatives}$ (11)

$JaccardIndex\left(JI\right)=\frac{TruePositives}{TruePositives+FalseNegatives+FalsePositives}$ (12)

Whereas, TP and TN relates that the tumor pixels and non-tumor pixels has been detected truly and FP and FN relates that the tumor pixels and non-tumor pixels has been detected falsely.

The parameters precision, recall and Jaccard index are measured in percentage with the aid of the manually tumor segmented images by the expert clinicians, is shown in equations 10, 11 and 12.

The proposed glioma detection approach stated in this research work are evaluated on the set of brain MRI images which are randomly selected from the TsM in KBI dataset and named as KBI1 to KBI10 and its evaluation experimental results are depicted in Table 3. The proposed MTDS approach obtains 99% of precision, 99% of Recall and 99.02% of JI in the set of 10 images in KBI dataset. The similar evaluation results are also obtained for testing with all the brain images in the TsM of the KBI dataset.

Table 3
Evaluation index estimation in KBI dataset images

KBI dataset image orders	Evaluation index in %
	Precision	Recall	JI
KBI1	98.8	99.3	99.3
KBI2	99.2	99.1	99.7
KBI3	98.9	98.7	99.2
KBI4	98.6	98.5	98.9
KBI5	99.3	99.1	98.5
KBI6	99.8	99.5	98.7
KBI7	99.1	99.2	99.1
KBI8	98.7	98.6	99.3
KBI9	98.4	98.9	98.6
KBI10	99.2	99.1	98.9
Average value in %	99	99	99.02

The proposed glioma detection approach stated in this research work are evaluated on the set of brain MRI images which are randomly selected from the TsM in BI20 dataset and named as BI1 to BI10 and its evaluation experimental results are depicted in Table 4. The proposed MTDS approach obtains 98.96% of precision, 98.91% of Recall and 98.78% of JI in the set of 10 images in KBI dataset. The similar evaluation results are also obtained for testing with all the brain images in the TsM of the BI dataset.

Table 4
Evaluation index estimation in BI20 dataset images

BI20 dataset image orders	Evaluation index in %
	Precision	Recall	JI
BI1	98.3	99.3	98.4
BI2	98.1	99.1	98.2
BI3	98.9	98.6	98.5
BI4	99.3	98.9	98.3
BI5	99.1	98.3	99.7
BI6	99.2	99.1	98.6
BI7	98.9	99.3	98.9
BI8	99.3	99.7	99.3
BI9	99.6	98.3	99.1
BI10	98.9	98.5	98.8
Average value in %	98.96	98.91	98.78

Table 5 is the evaluation index comparisons between datasets KBI and BI20 with respect to precision, recall and JI respectively.

Table 5
Evaluation index comparisons between datasets

Evaluation index in %	KBI dataset	BI20 dataset
Precision	99	98.96
Recall	99	98.91
JI	99.02	98.78

Figure 6
Evaluation index comparisons

Figure 6 shows the evaluation index comparisons between datasets KBI and BI20 with respect to precision, recall and JI respectively. Table 6 shows the comparisons of proposed MTDS approach with other existing approaches (KBI dataset). The existing methods Prakash and coauthors [²⁴] (2023) achieved 98.31% precision, 98.01% recall, 98.57% JI and 96.67% DR, Prakash BV and coauthors [²⁵] (2022) achieved 98.10% precision, 98.38% recall, 98.03% JI and 95.09% DR, Alqazzaz and coauthors [²⁶] (2022) achieved 97.38% precision, 97.17% recall, 97.19% JI and 95.10% DR, Amin and coauthors [⁶] (2022) achieved 96.57% precision, 96.46% recall, 96.97% JI and 94.84% DR, Tiwari and coauthors [²⁷] (2020) achieved 95.29% precision, 95.98% recall, 95.47% JI and 93.19% DR and Thiyaneswaran and coauthors [²⁸] (2020) achieved 94.98% precision, 94.29% recall, 94.97% JI and 94.86% DR [²⁴-²⁸].

Table 6
Comparisons of proposed MTDS approach with other existing approaches (KBI dataset)

Approaches	Precision(%)	Recall(%)	JI(%)	DR(%)
MTDS(in this research work)	99	99	99.02	98.95
Prakash et al. (2023)	98.31	98.01	98.57	96.67
Prakash BV et al. (2022)	98.10	98.38	98.03	95.09
Alqazzaz et al. (2022)	97.38	97.17	97.19	95.10
Amin et al. (2022)	96.57	96.46	96.97	94.84
Tiwari et al. (2020)	95.29	95.98	95.47	93.19
Thiyaneswaran et al. (2020)	94.98	94.29	94.97	94.86

Table 7 shows the comparisons of proposed MTDS approach with other existing approaches (BI20 dataset). The existing methods Prakash and coauthors [²⁴] (2023) achieved 98.16% precision, 97.56% recall, 97.37% JI and 97.46% DR, Prakash BV and coauthors [²⁵] (2022) achieved 97.59% precision, 97.16% recall, 97.85% JI and 96.09% DR, Alqazzaz and coauthors [²⁶] (2022) achieved 96.64% precision, 96.05% recall, 96.47% JI and 96.16% DR, Amin and coauthors [⁶] (2022) achieved 95.18% precision, 95.75% recall, 95.19% JI and 95.39% DR, Tiwari and coauthors [²⁷] (2020) achieved 94.84% precision, 94.47% recall, 94.37% JI and 93.29% DR and Thiyaneswaran and coauthors [²⁸] 2020) achieved 93.30% precision, 93.18% recall, 92.87% JI and 93.05% DR.

Table 7
Comparisons of proposed MTDS approach with other existing approaches (BI20 dataset)

Approaches	Precision(%)	Recall(%)	JI(%)	DR(%)
MTDS(in this research work)	98.96	98.91	98.78	99.2
Prakash et al. (2023)	98.16	97.56	97.37	97.46
Prakash BV et al. (2022)	97.59	97.16	97.85	96.09
Alqazzaz et al. (2022)	96.64	96.05	96.47	96.16
Amin et al. (2022)	95.18	95.75	95.19	95.39
Tiwari et al. (2020)	94.84	94.47	94.37	93.29
Thiyaneswaran et al. (2020)	93.30	93.18	92.87	93.05

Time complexity is an important evaluation parameter which estimates the computational time period for detecting the tumor or non-tumor image by the proposed approach stated in this research work. It is computed in milli seconds and this time complexity is based on the hardware utilizations. All existing approaches for brain tumor detection process used different hardware utilization and hence the time complexity cannot be compared. Hence, the hardware systems (Core i5 processor with 8 GB RAM) which are used in this research work are used in all the existing evaluation approaches. Table 8 is the time complexity comparisons between datasets.

Table 8
Time complexity comparisons between datasets

	Average Time consumption (ms) per image
Approaches	KBI dataset	BI20 dataset
MTDS(in this research work)	0.45	0.42
Prakash et al. (2023) [24]	0.94	0.91
Prakash BV et al. (2022) [25]	0.79	0.93
Alqazzaz et al. (2022) [26]	1.09	1.25
Amin et al. (2022) [6]	1.65	1.98
Tiwari et al. (2020) [27]	1.53	1.95
Thiyaneswaran et al. (2020) [28]	2.76	2.67

CONCLUSION

Using the suggested IVGG architecture along with a segmentation technique presented in this work, abnormal pixel regions in brain MRI images can be identified and segmented. Improved glioma detection rate is the goal of the IVGG, which is based on the current VGG design. The suggested MTDS effectively detected 210 glioma images in the KBI dataset, with a 99% Detection Ratio (DR), and 371 healthy images with a 98.9% DR, for an average DR of roughly 98.95%. Comparably, it found 718 healthy photos with a DR of 99.7% and 326 glioma images with a 98.7% DR in the BI20 dataset, for an average DR of roughly 99.2%. Furthermore, on both datasets, the suggested MTDS showed excellent precision, recall, and Jaccard Index (JI) percentages. Specifically, in the KBI dataset, it achieved 99% precision, 99% recall, and 99.02% JI, while in the BI20 dataset, it obtained 98.96% precision, 98.91% recall, and 98.78% JI. However, it's essential to note a limitation of this study: the proposed methods only detect tumor pixels without further diagnostic processing. To address this limitation, future work aims to expand the proposed approach by employing General Adversarial Networks (GAN) to not only detect but also diagnose tumor regions. Finally, the computational complexity of the proposed algorithm is evaluated and compared with existing works, highlighting its performance and potential areas for improvement.

REFERENCES

• ¹ Abd El Kader I, Xu G, Shuai Z, Saminu S, Javaid I, Ahmad IS, et al. Brain tumor detection and classification on MR images by a deep wavelet auto-encoder model. Diagnostics (Basel). 2021 Aug 31;11(9):1589. https://doi.org/10.3390/diagnostics11091589
  » https://doi.org/10.3390/diagnostics11091589
• ² Kshirsagar PR, Manoharan H, SivaNagaraju V, Alqahtani H, Noorulhasan Q, Islam S, et al. Accrual and dismemberment of brain tumours using fuzzy interface and grey textures for image disproportion. ComputIntellNeurosci. 2022 Jul 30;2022:2609387. https://doi.org/10.1155/2022/2609387
  » https://doi.org/10.1155/2022/2609387
• ³ Parameswari A, Bhavani S, Vinoth Kumar K. A convolutional deep neural network based brain tumor diagnosis using clustered image and feature-supported classifier (CIFC) technique. Braz Arch Biol Technol. 2023;66(4):1-15. https://doi.org/10.1590/1678-4324-2023230012
  » https://doi.org/10.1590/1678-4324-2023230012
• ⁴ Parameswari A, Vinoth Kumar K, Gopinath S. Thermal analysis of Alzheimer’s disease prediction using random forest classification model. Mater Today Proc. 2022;66(3):815-21. https://doi.org/10.1016/j.matpr.2022.04.357
  » https://doi.org/10.1016/j.matpr.2022.04.357
• ⁵ Kokila B, Devadharshini MS, Anitha A, Abisheak Sankar S. Brain tumor detection and classification using deep learning techniques based on MRI images. In: ICCCEBS 2021; 2022. doi: 10.1088/1742-6596/1916/1/012226
  » https://doi.org/10.1088/1742-6596/1916/1/012226
• ⁶ Amin J, Anjum MA, Sharif M, Jabeen S, Kadry S, Moreno Ger P. A new model for brain tumor detection using ensemble transfer learning and quantum variational classifier. ComputIntellNeurosci. 2022 Apr 14;2022:3236305. https://doi.org/10.1155/2022/3236305
  » https://doi.org/10.1155/2022/3236305
• ⁷ Prasad G, Gaddale VS, Kamath RC, Shekaranaik VJ, Pai SP. A study of dimensionality reduction in GLCM feature-based classification of machined surface images. Arab J Sci Eng. 2023; 49(4). http://dx.doi.org/10.1007/s13369-023-07854-1
  » http://dx.doi.org/10.1007/s13369-023-07854-1
• ⁸ Hephzipah JJ, Thirumurugan P. Performance analysis of meningioma brain tumor detection system using feature learning optimization and ANFIS classification method. IETE J Res. 2020;68(2):1542-50. https://doi.org/10.1080/03772063.2020.1844079
  » https://doi.org/10.1080/03772063.2020.1844079
• ⁹ Fayaz M, Torokeldiev N, Turdumamatov S, Qureshi MS, Qureshi MB, Gwak J. An efficient methodology for brain MRI classification based on DWT and convolutional neural network. Sensors (Basel). 2021 Nov 10;21(22):7480. https://doi.org/10.3390/s21227480
  » https://doi.org/10.3390/s21227480
• ¹⁰ Sasikanth S, Suresh Kumar S. Glioma tumor detection in brain MRI image using ANFIS-based normalized graph cut approach. Int J Imaging Syst Technol. 2017;28(1): 64-71. https://doi.org/10.1002/ima.22257
  » https://doi.org/10.1002/ima.22257
• ¹¹ Madhukumar S, Santhiyakumari N. Evaluation of k-means and fuzzy C-means segmentation of MR images of brain. J Nucl Med. 2015; 46(2):475-9. https://doi.org/10.1016/j.ejrnm.2015.02.008
  » https://doi.org/10.1016/j.ejrnm.2015.02.008
• ¹² Naser MA, Deen MJ. Brain tumor segmentation and grading of lower-grade glioma using deep learning in MRI images. Comput Biol Med. 2020;121:103758. https://doi.org/10.1016/j.compbiomed.2020.103758
  » https://doi.org/10.1016/j.compbiomed.2020.103758
• ¹³ Akter A, Nosheen N, Ahmed S, Hossain M, Yousuf MA, Abdullah Almoyad MA, et al. Robust clinical applicable CNN and U-Net based algorithm for MRI classification and segmentation for brain tumor. Expert Syst Appl. 2024;238:122347. https://doi.org/10.1016/j.eswa.2023.122347
  » https://doi.org/10.1016/j.eswa.2023.122347
• ¹⁴ Kurc T, Bakas S, Saltz J, Manjón JV, Kalpathy-Cramer J, Flanders A, et al. Segmentation and classification in digital pathology for glioma research: Challenges and deep learning approaches. Front Neurosci. 2020;14:27. doi: 10.3389/fnins.2020.00027.
  » https://doi.org/10.3389/fnins.2020.00027.
• ¹⁵ Aurna NF, Yousuf MA, Taher KA, Azad AKM, Moni MA. A classification of MRI brain tumor based on two stage feature level ensemble of deep CNN models. Comput Biol Med. 2022 Jul;146:105539. https://doi.org/10.1016/j.compbiomed.2022.105539
  » https://doi.org/10.1016/j.compbiomed.2022.105539
• ¹⁶ Díaz-Pernas FJ, Martínez-Zarzuela M, Antón-Rodríguez M, González-Ortega D. A deep learning approach for brain tumor classification and segmentation using a multiscale convolutional neural network. Healthcare (Basel). 2021 Feb 2;9(2):153. https://doi.org/10.3390/healthcare9020153
  » https://doi.org/10.3390/healthcare9020153
• ¹⁷ Aurna NF, Yousuf MA, Taher KA, Azad AKM, Moni MA. Multi-Classification of Brain Tumors via Feature Level Ensemble of Convolutional Neural Network. In: 2021 3rd Int Conf Sustain Technol Ind 4.0 (STI). 2021; 1-6. https://doi.org/10.1109/STI53101.2021.9732543
  » https://doi.org/10.1109/STI53101.2021.9732543
• ¹⁸ Ahmad M, Ai D, Furqan Qadri S, Xie G. Deep belief network modeling for automatic liver segmentation. IEEE Access. 2019;7:20585-95. https://doi.org/10.1109/ACCESS.2019.2896961
  » https://doi.org/10.1109/ACCESS.2019.2896961
• ¹⁹ Bhavani R, Vasanth K. Brain image fusion-based tumour detection using grey level co-occurrence matrix Tamura feature extraction with backpropagation network classification. Math Biosci Eng. 2023 Mar 7;20(5):8727-44. https://doi.org/10.3934/mbe.2023383
  » https://doi.org/10.3934/mbe.2023383
• ²⁰ Shamna NV, Aziz Musthafa B. Feature extraction method using HoG with LTP for content-based medical image retrieval. Int J ElectrComput Eng Syst. 2023;14(3):267-75. https://doi.org/10.32985/ijeces.14.3.4
  » https://doi.org/10.32985/ijeces.14.3.4
• ²¹ Parameswari A, Bhavani S, Vinoth Kumar K. A deep learning based glioma tumour detection using efficient visual geometry group convolutional neural networks architecture. Braz Arch Biol Technol. 2024;67.https://doi.org/10.1590/1678-4324-2024230705
  » https://doi.org/10.1590/1678-4324-2024230705
• ²² Anita JN, Kumaran S. A deep learning architecture for meningioma brain tumor detection and segmentation. J Cancer Prev. 2022 Sep 30;27(3):192-8. https://doi.org/10.15430/jcp.2022.27.3.192
  » https://doi.org/10.15430/jcp.2022.27.3.192
• ²³ Arash R, Masoud B, Masoumeh G. Automated segmentation and morphological characterization of placental intervillous space based on a single labeled image. Micron. 2023;169:103448. https://doi.org/10.1016/j.micron.2023.103448
  » https://doi.org/10.1016/j.micron.2023.103448
• ²⁴ Prakash M, Neelakandan S, Tamilselvi M, Velmurugan S, Baghavathi Priya S, Martinson EO. Deep learning-based wildfire image detection and classification systems for controlling biomass. Int J Intell Syst. 2023:1-18. https://doi.org/10.1155/2023/7939516
  » https://doi.org/10.1155/2023/7939516
• ²⁵ Prakash BV, Kannan AR, Santhiyakumari N, Kumarganesh S, Raja DSS, Hephzipah JJ, et al. Meningioma brain tumor detection and classification using hybrid CNN method and RIDGELET transform. Sci Rep. 2023 Sep 4;13(1):14522.https://doi.org/10.1038/s41598-023-41576-6
  » https://doi.org/10.1038/s41598-023-41576-6
• ²⁶ Alqazzaz S, Sun X, Nokes LD, Smith S, Morabito D, Wu L. Combined features in region of interest for brain tumor segmentation. J Digit Imaging. 2022;35(6):938-46. doi: 10.1007/s10278-022-00602-6.
  » https://doi.org/10.1007/s10278-022-00602-6.
• ²⁷ Tiwari A, Srivastava S, Pant M. Brain tumor segmentation and classification from magnetic resonance images: Review of selected methods from 2014 to 2019. Pattern Recognit Lett. 2020;131:244-60. https://doi.org/10.1016/j.patrec.2019.11.020
  » https://doi.org/10.1016/j.patrec.2019.11.020
• ²⁸ Thiyaneswaran K, Anguraj S, Kumarganesh K, Thangaraj. Early detection of melanoma images using gray level co-occurrence matrix features and machine learning techniques for effective clinical diagnosis. Int J Imaging Syst Technol. 2020;31(2):682-94. http://dx.doi.org/10.1002/ima.22514
  » http://dx.doi.org/10.1002/ima.22514

• Funding:
  The authors declare they have no funding applied.

Edited by

• Editor-in-Chief:
  Alexandre Rasi Aoki

• Associate Editor:
  Raja Soosaimarian Peter Raj

Publication Dates

• Publication in this collection
  03 Feb 2025
• Date of issue
  2025

History

• Received
  05 Feb 2024
• Accepted
  29 Aug 2024
• Corrected
  05 Feb 2025