1.
Introduction
The human brain, the most intricate organ in the human body, orchestrates an array of vital biological functions, including respiration and muscular control. However, when a mass of cells undergoes abnormal growth within or adjacent to the brain, it culminates in a formidable challenge—brain tumors. Among these, gliomas [1] emerge as one of the most common categories of glial-based primary tumors in adults. They exhibit varying degrees of malignancy, encompassing Grade Ⅰ to Grade Ⅳ, as classified by World Health Organization [2,3]. Grades Ⅰ and Ⅱ are typically deemed benign or low-grade (LGG), while Grades Ⅲ and Ⅳ are classified as malignant or high-grade (HGG) [4]. Grade Ⅰ tumors grow slowly and are sometimes entirely resectable with surgery, but Grade Ⅳ tumors are aggressive, rapidly growing, and challenging to treat. The most frequent primary tumors were astrocytomas (38.7%), with HGGs (59.5%) making up the majority [4]. Despite recent advancements in therapeutic and diagnostic techniques for malignant gliomas, the median survival for patients with Grade Ⅲ is less than five years, and for those diagnosed with Grade Ⅳ, it stands at a mere 15 months [5]. Table 1 presents various glioma grades categorized by their malignancy and level of aggressiveness. Neuroimaging data, particularly Magnetic Resonance Imaging (MRI) sequences such as T1C, T1W, T2W, and FLAIR (Fluid-Attenuated Inversion Recovery), offer precious insights into the size, shape, location, and metabolic activity of brain tumors [6]. These data serve as a crucial foundation for clinicians to assess the tumor's condition both before and after therapeutic interventions. The wealth of information contained in these images can be harnessed to distinguish between various brain tissues, facilitating the discrimination between tumor and normal [7,8,9].
Efficient management of gliomas necessitates two fundamental steps: the precise segmentation of the tumor to locate and characterize its various components and the accurate classification of the tumor to determine its grade. However, the conventional approach to these tasks involves manual intervention by physicians, which is not only laborious and time-consuming but also susceptible to inter-observer variability among specialists, thus potentially impeding the efficacy of treatment [10,11].
Considering these challenges, the quest for automated diagnosis systems has gained momentum, aiming to equip clinicians with timely and precise morphological information about different tumor regions. This information is pivotal for facilitating prompt and accurate tumor diagnosis and treatment [11]. Recent decades have witnessed significant strides in medical image analysis, with machine learning (ML) emerging as a promising avenue. Traditional ML approaches, while effective, often entail the prerequisite of prior knowledge and manual feature extraction, which can be time-consuming [12]. In contrast, deep learning (DL), a more recent subfield of ML, has exhibited notable advantages in addressing the limitations of traditional ML based models. DL models can automatically extract intricate high-level features and seamlessly integrate feature extraction with classification/segmentation, streamlining the overall process [13,14]. Among the various DL approaches, Convolutional Neural Network (CNN) architectures have demonstrated superior performance, especially in the detection and analysis of neurological diseases [15]. To substantiate the viability of CNNs, they autonomously generate robust discriminative features through a hierarchical learning methodology. In the initial layers, feature maps extract fundamental low-level characteristics, while in subsequent layers, they capture more sophisticated, domain-specific attributes. The feature maps in the early layers encode basic structural elements such as edges, shapes, and textures, while the higher layers amalgamate these lower-level feature maps to construct abstract representations that encompass both local and global information. As a result, DL transforms manual feature extraction into a self-learning process. CNNs have been in existence for many years, but their popularity surged when Krizhevsky et al. [16] pioneered the use of DL with AlexNet. Simonyan and Zisserman [17] later introduced a similarly deep CNN, VGG Net, excelling in localization and classification tasks. The success of DL, particularly CNNs, can be attributed to advancements in computational technologies, such as powerful GPUs, and the evolution of learning algorithms. Some of the DL-based research efforts in glioma grading are discussed below.
Hao et al. [18] introduced a fully automated brain tumor detection and segmentation approach based on U-Net deep convolution networks, achieving efficient and robust segmentation results compared to manual ground truth delineation. Khawaldeh et al. [19] proposed the use of the AlexNet DL algorithm for identifying and grading abnormal brain MRIs; however, this method exhibited limitations in handling small datasets. Anaraki et al. [20] introduced gaussian assumption to streamline the selection of deep neural network architectures, eliminating the need for trial and error. They also employed an ensemble algorithm to reduce prediction errors, achieving an impressive 90.9% accuracy in classifying gliomas into three grades. Mzoughi et al. [21] leveraged volumetric T1-Gado MRI sequences to develop an automatic deep multi-scale 3D CNN model for glioma classification, achieving an overall accuracy of 96.49%.Zhuge et al. [22] utilized a modified U-Net model for segmentation and a 3D volumetric CNNs (3DConvNet) classification model to grade gliomas, attaining an accuracy of 97.1%.Gutta et al. [23] constructed a DCNN model for glioma classification and demonstrated its superiority over traditional ML models with an accuracy of 87%.Lu et al. [24] employed the ResNet model for glioma classification and incorporated pyramid dilated convolution to enhance performance; however, this method was limited to 2D MRI interpretation and required manual labeling of the training set. In a study by Naser et al. [25], a multi-task DL-based method was proposed, utilizing the U-Net architecture for segmentation and VGG-16 for classification. The reported dice score and tumor detection accuracy were 0.84 and 0.92, respectively. Decuyper et al. [26] trained a pipeline consisting of U-Net for segmentation and ResNet for classification, achieving an accuracy rate of 90% and a sensitivity rate of approximately 93.48%. Furthermore, Tandel et al. [27] developed four therapeutically relevant datasets and employed five DL-based models (AlexNet, VGG-16, ResNet18, GoogleNet, and ResNet50) along with five ML-based models. They introduced the MajVot method to maximize classification performance and reported an improvement in average accuracy. In their subsequent work [28], they worked with five pre-trained CNNs, achieving a maximum accuracy of 99.06% through a majority voting technique. Nassar et al. [29] also used the concept of majority voting to the prediction outputs of five different CNN models (GoogleNet, AlexNet, SqueezeNet, ShuffleNet, and NasNet-Mobile) to get an accurate classification of brain tumors based on the T1c MRI dataset, which achieved an accuracy of 96.08%, 95.16%, 96.67%, 97.17% and 97.5% respectively.
From the brief literature review, it could be concluded that TL and ensemble learning emerge as the two predominant research-driven and application-oriented subfields in the realm of glioma classification and grading methods using DL. TL [30] leverages pre-trained neural network models, fine-tuning them to adapt to the specific characteristics of glioma data, thus capitalizing on the knowledge gained from other domains. Ensemble learning [31], on the other hand, encompasses a range of techniques that combine the predictions of multiple models to improve the overall accuracy and robustness of the classification and grading processes. These two subfields represent the forefront of research efforts in the quest for more accurate and clinically relevant solutions in the field of glioma analysis. Apart from this, the major drawbacks in most of the reviewed approaches for glioma classification and grading are summarized as follows:
● DL requires large, well-annotated datasets, which are often lacking in glioma research.
● Ensuring models perform well on diverse patient data is an ongoing challenge.
● Fine-tuning complex models demands substantial computational resources and expertise.
● Lack of standardized pre-processing affects data quality.
● Limited research has compared models with and without segmentation to evaluate the impact, and no studies provide clear statistical evidence for directly incorporating segmentation results into classification models.
The motivation for this research stems from the pressing need for more accurate, automated, and efficient methods in glioma diagnosis and grading. The complexity of gliomas, with their varied shapes, sizes, and locations within the brain, demands advanced image analysis techniques for accurate delineation. While advanced medical imaging techniques like MRI scans have enhanced our ability to visualize and diagnose brain tumors, manual analysis remains time-consuming and susceptible to human error. DL and image processing offer the potential to automate this process, potentially improving diagnostic accuracy and treatment planning reliability. In recent years several studies have explored various techniques, achieving promising results in tumor segmentation, classification, and even survival prediction. Although previous research has explored both segmentation and classification individually, fewer studies have focused on integrating them for a more comprehensive approach. Our novel contribution addresses this gap by integrating customized segmentation and classification architectures. Specifically, we have tailored the U-Net model for segmentation and the VGG-16 model for classification to better suit our research objectives. Our research leverages the BraTS2018 dataset, a widely recognized benchmark in the field of neuroimaging [32,33,34], to ensure the reliability and relevance of our findings. The methodology begins with preprocessing steps, including data splitting and application of gamma correction for enhanced image detail, and normalization to standardize pixel values. Rotation augmentation is implemented to improve the model's generalization ability. Segmentation is performed using U-Net, featuring a tailored encoder-decoder, hybrid loss function and Leaky ReLU activation for improved gradient stability. For classification, we leverage fine-tuned VGG-16 model, incorporating strategic layer freezing for optimal performance. Both the segmentation and classification models undergo systematically tuned training, ensuring robust and accurate analysis throughout the study. While the novelty of our approach lies in the integration of these modified segmentation and classification architectures, this unique combination not only enables precise tumor localization but also provides a means to categorize gliomas into low-grade and high-grade groups, a critical factor in treatment planning. Leveraging this combined approach, our research aims to substantially contribute to ongoing efforts to improve glioma diagnosis and, consequently, enhance treatment outcomes by harnessing the capabilities of DL in medical imaging.
The paper is structured as follows: Section 2 introduces the proposed methodology, Section 3 showcases experimental results and analysis, Section 4 is dedicated to discussion and limitations, and finally, Section 5 concludes the research and provides future directions.
2.
Materials and methods
2.1. Dataset
The proposed scheme is evaluated on BRaTS 2018 dataset obtained from Multimodal Brain Tumor Segmentation Challenge 2018. This dataset comprises 285 cases, including 210 High-Grade Gliomas (HGGs) and 75 Low-Grade Gliomas (LGGs). The dataset has been split into training, validation, and test subsets, and within the validation subset, there are 66 glioma tumors of various grades. The training set was further divided into 80% for training and 20% for testing, resulting in a total of 228 training images. The images are provided in NIfTI (Neuroimaging Informatics Technology Initiative) format, covering four modalities: T1W, T1C, T2W, and FLAIR. All the images within this dataset share a uniform volume dimension of 240×240×155, and they have been subject to careful manual annotation conducted by expert neuro-radiologists. This annotation aimed to delineate four distinct intra-tumoral regions, which include the necrotic core, non-enhancing, edema, and enhancing tumor areas. The dataset encompasses diverse glioma types with varying shapes, sizes, and locations within the brain. Each MRI sequence captures unique brain features using different scales, unlike standard images confined to a 0-255 gray value range. This variability stems from heterogeneous origin of the dataset, with scans acquired from various MRI scanners across different institutions. To address this inconsistency and ensure efficient training and inference, we normalize the voxel intensities using Equation 3. The schematic representation depicting the procedural framework employed in our study is illustrated in Figure 1.
2.2. Pre-processing
In the data preprocessing phase, the BraTS2018 dataset was prepared to ensure its suitability for training and evaluating the DL models. All scans were saved in neuroimaging format (i.e. nii.gz), so the first step is to convert nii.gz to 3D NumPy array. The use of gamma correction [35] was applied to address variations in image illumination and enhance the visibility of subtle structures within the MR images. Mathematically, this correction can be represented as follows:
where, γ is the correction factor.
Gamma correction was applied to enhance MRI modalities, except for FLAIR images. Gamma values of 2.3, 2.5, and 3 were chosen for T1W, T1C, and T2W modalities, respectively, to improve segmentation performance. This modality-specific gamma correction tailors contrast enhancement to each MRI type thereby optimizing the visibility of different tissue types and pathological features unique to each modality.
Additionally, data augmentation techniques, such as rotation, were employed to increase the diversity of the training data and improve the model's generalization ability [36]. Rotation augmentation helps the model become invariant to different orientations of gliomas within the images. This rotation operation can be expressed mathematically as:
where, θ represent the rotation angle.
The value of θ ranging from 0 to 30 degrees are chosen randomly along one of the three dimensions of the 3D pictures to simulate variations in glioma orientation within the brain, ensuring that the model can handle tumors at different angles and positions. This range is chosen to introduce variability in the dataset without drastically altering the original image characteristics. Rotations within this range can simulate different patient orientations during MRI scans, making the model more robust to such variations in real-world scenarios. However, larger rotations might distort the anatomical structures in the images, leading to unrealistic scenarios and potentially misleading the model.
Furthermore, to ensure consistency in model input sizes and reduce computational complexity, the MRI images were cropped to a uniform size of 190×190×155 pixels. This size reduction was based on considerations of memory constraints and computational efficiency while still preserving relevant anatomical information. Finally, data normalization was performed on each sequence of MRI images using the following equation:
where, μ is the mean intensity value within the image, and σ is the standard deviation of the intensities within the image slices. Each MRI sequence is individually normalized using its own mean and standard deviation, instead of using dataset-wide values. This approach aims to improve generalizability while addressing the diverse intensity ranges encountered in our dataset, potentially leading to better model performance compared to standard normalization process. This normalization ensures that the input data are cantered and scaled appropriately for the DL models [37].
2.3. Segmentation
In this study, U-Net [38] is employed for glioma segmentation because of its exceptional architecture, featuring contracting and expansive pathways, which efficiently capture intricate anatomical structures and tumor boundaries in MR imaging data. Figure 2 illustrates the proposed model. Our network comprises five convolutional blocks, incorporated within both the encoder and decoder paths. These convolutional layers employ a 3x3 kernel size, facilitating the extraction of intricate spatial features while maintaining computational efficiency. A stride of 1 is consistently applied in both horizontal and vertical directions, ensuring meticulous convolutional operations across the input. Furthermore, a 2x2 maxpooling size is employed to down sample feature maps while retaining essential information. The activation function chosen is Leaky ReLU, for its capacity to mitigate the vanishing gradient problem [39]. For the final layer, the network adopts a softmax activation function, essential for pixel-wise classification in multi-class segmentation. Incorporating a dropout rate of 0.2 bolsters the network's robustness against overfitting while maintaining model expressiveness. The hybrid loss function employed in this study is defined by the following equation:
Here, the weights α and β were set to 0.3 and 0.6, respectively, through a rigorous grid search process to find the optimal balance for our dataset. The α coefficient scales the contribution of the categorical cross entropy component, which is a standard choice for multi-class classification tasks [40]. However, it may not fully capture the desired segmentation quality, especially for imbalanced classes. To address this, we combine categorical cross entropy with jaccard index. Jaccard index measures the overlap between predicted and ground truth tumor regions, providing a good understanding of segmentation accuracy. The β coefficient is used to adjust the importance of the Jaccard similarity-based term, (1 − J_tumor), which is designed to encourage the model to focus on accurate segmentation of the tumor region. This loss function is a critical element of the network's training process, aiding in achieving the desired balance between classification accuracy and precise tumor region delineation. In terms of training, the network undergoes a learning rate of 0.01 and is trained for 8 epochs, the optimal settings established through iterative experimentation. For parameter optimization, the study employs the stochastic gradient descent (SGD) algorithm with a momentum of 0.99. The proposed model differs from the original U-Net by including a novel hybrid loss function, adopting Leaky ReLU activation to improve gradient stability, and implementing dropout regularization.
2.4. Classification
VGG-16 [17] is an established CNN architecture comprising a total of 16 layers organized into five convolutional blocks, with Leaky ReLU activation functions applied in each block. Following each block, max-pooling layers are applied, and the architecture is concluded with three fully connected layers, as illustrated in Figure 3. In pre-trained CNNs, the lower layers capture basic, general features, while the upper layers specialize in more intricate features. Consequently, we typically freeze the parameters of the earlier layers to maintain their focus on low-level features. In our current study, only the last three layers were kept trainable, primarily for classification purposes. For optimization, we utilized the SGD algorithm and employed categorical cross-entropy as the loss function [41]. Furthermore, we set specific hyperparameters, including a learning rate of 0.001, a batch size of 16, and ran training for 8 epochs The model's construction was executed using Keras, with TensorFlow serving as the backend framework. The novelty of our architecture, based on VGG-16, lies in its adaptation for glioma classification: transfer learning from a general domain to medical imaging, a customized output layer, strategic layer freezing, tailored hyperparameters, and the use of dropout for regularization. These adaptations enhance its suitability for the specific medical image classification task.
3.
Results
In this section, we present the results of glioma segmentation and classification using U-Net and VGG-16 models, followed by a comparative analysis with state-of-the-art methods. The proposed model, implemented using Python 3.7, TensorFlow 2.0, and Keras 2.3.1 frameworks, was evaluated on the BraTS 2018 dataset. Experiments were conducted on an Nvidia GeForce RTX 2080 Ti GPU with 11GB RAM. The initial preprocessing involved converting NIfTI images to 3D NumPy arrays and applying gamma correction, rotation augmentation, cropping, and normalization to enhance model robustness. For segmentation, we utilized a U-Net architecture featuring five convolutional blocks within both the encoder and decoder paths. Leaky ReLU activation, a hybrid loss function, and dropout regularization were incorporated to ensure accurate delineation of tumor regions. Classification leveraged a modified VGG-16 architecture with strategic layer freezing, tailored hyperparameters, and dropout. The final layers of the fine-tuned CNN models were adapted to accommodate the two tumor classes (LGG and HGG). Model optimization included a comprehensive grid search, optimizing key parameters like learning rate and dropout rate. Empirical observations revealed an inversely proportional relationship between learning rate/batch size and performance/training time. Consequently, smaller learning rates and batch sizes were chosen to improve performance. Model evaluation was conducted on the validation set using metrics: categorical cross entropy loss, jaccard index, and accuracy. Early stopping based on validation loss prevented overfitting, and robust 5-fold cross-validation ensured generalizability by continuously monitoring metrics within each fold.
3.1. Performance Evaluation Metrics
This study aims to determine the class of glioma tumors using MR images, which can determine whether a patient had LGG or HGG. Depending on the result discovered, output can be either positive or negative. The dice score [42] serves as a widely utilized validation metric for medical image segmentation, facilitating comparisons between automatic segmentation outcomes and ground truth, while the jaccard index offers another valuable measure for assessing the overlap and accuracy of segmentation results. Additionally, it assesses the reproducibility of segmentation. Accuracy, a key efficiency metric in the classification process, is employed when equal importance is assigned to both positive and negative classes. Recall measures the proportion of correctly predicted positive instances, while specificity identifies the percentage of accurately recognized true negatives [43]. Table 2 presents the metrics used to evaluate the model's effectiveness.
3.2. Glioma Segmentation with U-Net
Our utilization of the U-Net model for glioma segmentation yields compelling results, underscoring its efficacy in precisely delineating glioma regions within MR images. Table 3 summarizes glioma segmentation results using the U-Net model, showcasing high accuracy. The Dice Score values for whole tumor (WT), tumor core (TC) and enhanced tumor (ET) are 0.96, 0.92, and 0.89, respectively. The jaccard index values are 0.92, 0.85, and 0.80, and Recall values are 0.89, 0.94, and 0.87, indicating robust segmentation performance across different glioma regions. These results exhibit a substantial alignment with the ground truth.
3.3. Glioma Classification with VGG-16
We further evaluate the performance of our model by classifying gliomas into LGGs and HGGs categories using the VGG-16 model. Classification results are presented in Table 4. These results offer valuable insights into the performance of our grading model on the BraTS 2018 dataset. The model achieved an overall accuracy of 97.89%, indicating its ability to correctly classify glioma grades. All metrics exceeding 0.95 for both categories underscore the model's effectiveness in distinguishing between both grades. Figure 4(a) illustrates our model's ability to distinguish between LGG and HGG through a Receiver Operating Characteristic (ROC) curve. A higher curve and greater area under it indicate better classification accuracy. The comparison of the model's predictions against the actual ground truth labels for the BraTS 2018 dataset is illustrated in Figure 4(b).
3.4. Comparison with State-of-the-Art
Table 5 presents a comprehensive comparative study of glioma segmentation methods, referencing various approaches and evaluating their performance. It investigates diverse DL based architectures and MRI modalities., Our model demonstrates superior performance compared to established methods, achieving remarkable dice scores of 0.96, 0.92, and 0.89 for WT, TC, and ET, respectively. Figure 5(a) complements Table 5 by visually representing the dice scores, enabling an at-a-glance comparison of the segmentation approaches. Figure 5(b) presents a visual comparison of the segmentation result and ground truth from a validation set sample image.
Table 6 provides a comprehensive comparative analysis of various glioma classification methods, shedding light on their performance across different datasets, MRI modalities, architectures, and validation techniques. Our proposed model achieves competitive accuracy in a 5-fold cross-validation on the BraTS 2018 dataset. To provide a visual representation of these accuracy comparisons among the models mentioned in the table, Figure 6 visually summarizes their performance. Table 7 provides a detailed comparison of different approaches, shedding light on their efficacy in improving the precision and accuracy of glioma grading by integrating DL based segmentation before grading. It highlights the potential benefits of integrating segmentation and classification, exemplified in the table, as they lead to enhanced precision and accuracy in grading which is crucial for guiding clinical decisions and improving patient care.
4.
Discussion
In the existing body of literature, the interplay between segmentation and classification in the context of glioma analysis has been largely unexplored. Most studies have treated these as separate processes, focusing either on the segmentation of gliomas or on their classification into LGGs and HGGs. However, our approach addresses this gap by integrating segmentation and classification into a single, cohesive workflow. In this study, we explore the effectiveness of DCNNs using the tailored U-Net architecture for segmentation and the VGG-16 model for classification. Our primary aim is to simultaneously tackle glioma segmentation and classify gliomas into LGG and HGG categories using MRI data. To establish consistency and reliability in data pre-processing, we introduce a standardized methodology to ensures that our experiments are underpinned by a consistent pre-processing pipeline, ultimately enhancing data quality, and facilitating reproducibility. We strategically selected U-Net for segmentation and VGG-16 for classification, leveraging their complementary strengths. U-Net excels in capturing intricate anatomical structures and precise tumor boundaries, making it well-suited for detailed segmentation. Its unique architecture, incorporating skip connections, ensures the preservation of fine details crucial for accurate localization. The efficiency of U-Net in handling limited data and its end-to-end learning approach simplify the segmentation workflow. On the other hand, VGG-16, with its pre-trained layers, offers powerful feature extraction capabilities, enhancing computational efficiency through TL. It focuses on global patterns and class discrimination, making it ideal for classifying tumors into different subtypes based on overall characteristics. By combining U-Net's segmentation precision with VGG-16's feature extraction and classification abilities, we create a complementary workflow for comprehensive glioma analysis. This synergistic approach effectively addresses both the intricate details of tumor segmentation and the accurate classification of glioma grades.
Our U-Net-based glioma segmentation demonstrates exceptional precision, as indicated by dice score values of 0.96, 0.92, and 0.89 for WT, TC, and ET, as detailed in Table 3. These results underscore the accuracy and reliability of our model in delineating glioma regions within MRIs. The robust jaccard index further validates the accuracy of our model's spatial predictions compared to ground truth. A distinguishing aspect of our research lies in the integration of these customized segmentation and classification architectures. Our fine-tuned VGG-16 classification model achieves an overall accuracy of 97.89%, complemented by high specificity and recall values exceeding 0.95 for both LGGs and HGGs, as presented in Table 4. These results demonstrate proficiency of our model in differentiating between these clinically significant categories.
To further emphasize the novelty and significance of our work, Table 5 provides a comprehensive comparative investigation into glioma segmentation methodologies, assessing a range of approaches, and evaluating the performance of the proposed model against state-of-the-art methods. Notably, most of these methods leverage U-Net as their foundation. This demonstrates the impact of U-Net, as its versatility has led to numerous modifications and extensions for diverse medical image segmentation tasks. Mahasin et al. [60] achieved 95% accuracy in head MRI tumor detection with both U-Net and UBNet, demonstrating its continued relevance. Similar successes span other modalities: Sateesh et al. [61] utilized super-resolution and federated TL for lung segmentation, while Ryu et al. [62] proposed SegR-Net for retinal vessel segmentation. Further, Tiwari et al. [63] and Ashwini et al. [64] leveraged modifications like inception modules and improved training strategies. Beyond U-Net, alternative methodologies have demonstrated notable accuracy. For instance, Ranjbarzadeh et al. [65] introduced a segmentation framework that integrates feature selection via the improved chimp optimization algorithm alongside hyperparameter optimization. In another study Ranjbarzadeh et al. [66] proposed a cascaded CNN model for mammogram classification. This model used 11 encoded images derived from the original mammogram via a GLCM-based encoding algorithm, allowing extraction of important features for breast and tumor tissue boundary detection. Additionally, Kasgari et al. [67] achieved notable performance by leveraging zernike moments and LDNP encoding, surpassing several conventional CNN approaches. Many more similar studies highlight the great impact of DL in the medical field [68]. Notably, while many of these studies achieved promising brain tumor segmentation results on BraTS, none addressed the crucial aspect of treatment response assessment through combined segmentation and classification, which remains our key contribution. Table 6 delivers an in-depth comparative analysis of glioma classification strategies, highlighting their performance across a diverse array of DL-based architectural frameworks. To facilitate benchmarking, we have compiled results from several relevant studies. Our proposed model achieved 97.89% accuracy, exceeding most reported results [51,52,53]. While Hamdaoui et al. [54] and Khazaee et al. [55] showed slightly higher accuracy, their studies focused solely on classification for treatment response assessment. Table 7 presents a comparative study for glioma grading that integrates both segmentation and classification. This table emphasizes our innovative approach and its performance, showcasing our method's superiority with other state-of-the-art techniques. Most significantly, this approach delves into the underexplored area of the impact of segmentation on classification, effectively bridging a notable gap in the existing literature. Our research contributes empirical evidence regarding the influence of segmentation on classification, adding depth to the understanding of glioma diagnosis.
Despite these achievements, it is crucial to acknowledge the limitations inherent in this study, particularly our reliance on data from a single institution. Future research endeavors should consider the inclusion of data from multiple institutions to enable more extensive clinical trials and fortify the generalizability of our findings. Additionally, it is important to recognize that model performance may exhibit variations when applied to larger and more diverse datasets. Further refinement of the model architectures and hyperparameter tuning may lead to overall performance improvement, and these avenues represent potential future research directions.
While CNNs have dominated medical image segmentation, transformers are emerging as powerful alternatives. Many recent works have leverage transformers for improved accuracy and efficiency in medical image segmentation. Zhu et al. [69] propose a swin transformer-based method for brain tumor segmentation, combining it with CNNs for edge detection and feature fusion. Li et al. [70] introduce X-Net, a novel architecture combining CNNs and transformers for enhanced local and global feature extraction. Xianyu et al. [71] present a cloud-based segmentation method using transformers and CNNs, overcoming limitations in local computing power. Yang et al. [72] propose a hybrid network integrating CNNs and transformers with a multi-dimensional statistical feature extraction module for improved texture feature extraction. Building upon these works, future studies could investigate novel architectures that leverage the strengths of transformers, potentially combining them with CNNs or other techniques for even more accurate and efficient segmentation. Such advancements hold immense potential for further refining glioma diagnosis.
5.
Conclusion
In this research, we have presented a comprehensive approach for glioma segmentation and classification using DL models, specifically U-Net for segmentation and VGG-16 for classification. Our primary goal is to identify the most optimum channel for glioma grading and assess the significance of segmentation in classification models. Glioma segmentation with U-Net achieved high DS values of 0.96, 0.92 and 0.89 for WT, TC, and ET respectively, on the test dataset, demonstrating the model's ability to accurately outline tumor regions in MRI images. Furthermore, the glioma classification using the fine-tuned VGG-16 model resulted in an impressive accuracy of 97.89%, with consistently high specificity and recall values above 0.95 for both grades. These results highlight the model's ability to distinguish between different glioma grades. The synergy between U-Net and VGG-16 adds a layer of depth to our methodology, resulting in improved accuracy in glioma grading.
This study underscores the potential utility of DL-based glioma segmentation and classification in the field of medical imaging. By demonstrating the effectiveness of this integrated approach, our study provides empirical evidence supporting the hypothesis that accurate segmentation can indeed enhance the performance of subsequent classification. This finding not only validates our methodology but also opens new avenues for future research in this domain. We believe that our work will inspire further investigations into the synergistic effects of segmentation and classification, ultimately leading to more accurate and reliable glioma diagnosis. Future research should focus on integrating additional clinical data and developing real-time diagnostic tools for practical use in clinical settings. Additionally, broader clinical trials involving data from multiple institutions would enhance the generalizability of our findings.
Use of AI tools declaration
The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.
Acknowledgments
We would like to thank Department of Computer Science and Engineering, National Institute of Technology Agartala, Tripura, India for their support related to infrastructure. No funds gathered by any of the sources.
Conflict of interest
The authors declare there is no conflict of interest.
DownLoad:
