Brain tumor segmentation is an essential process in medical image analysis, which aims to pinpoint the affected areas of the brain due to the presence of a tumor [1]. Diagnosis, therapy planning, disease progression monitoring, and precise and effective segmentation of brain tumors are crucial [2]. The complex nature of brain tumors and the differences between patients make manually identifying these tumors a tough and time-consuming job [3]. Brain tumors represent a heterogeneous group of intracranial neoplasms that affect both adults and children, posing significant challenges for diagnosis and treatment [4]. Magnetic resonance imaging (MRI) stands as the top choice for non-invasive brain tumor detection and assessment because of its exceptional resolution and outstanding contrast for soft tissues [5]. In many clinical tasks—for example, diagnosis, treatment planning, and patient monitoring—the accurate segmentation of brain tumors using MRI data is crucial [6]. The traditional segmentation methods are mainly based on handcrafted features, and these are designed based on domain knowledge [7]. The problem is that they are generally sensitive to variations in image intensity and hence require extensive manual tuning. Thus, the robustness and precision are very low [8]. Segmentation of brain tumor techniques is crucial for accurate diagnosis, monitoring of tumor progression, and treatment planning [9]. These techniques can be generally divided into three categories: manual brain tumor segmentation, semi-automatic, and fully automatic methods [10]. Manual segmentation of the brain is performed by radiologists or experts and involves the delineation of tumor regions on medical images using graphical tools [11]. This method can be accurate, but it is slow and takes a lot of work [12]. Furthermore, the increasing demand for medical imaging and the limited availability of expert radiologists make manual segmentation challenging to scale. Semi-automatic methods require minimal user intervention, often providing an initial contour or seed point to guide the segmentation process. These methods rely on algorithms such as region growing, which iteratively groups neighboring pixels with similar intensity values [13]; level-set methods, which evolve a contour based on geometric and image-based properties [14]; and active contours or snakes, which deform a curve or surface to minimize an energy function derived from image features [15]. Semi-automatic methods offer improved efficiency compared to manual methods; however, they still require user interaction, which can be time-consuming and may introduce variability. In fully automatic methods, tumors can be segmented without user interaction, such as machine learning (ML) and DL approaches. These techniques seek to increase the segmentation process’ effectiveness, consistency, and scalability [16]. Handcrafted feature-based methods involve extracting engineered features from images and training ML classifiers for tumor segmentation, while DL techniques such as CNN automatically learn hierarchical representations of the data to perform segmentation [17]. Fully automatic methods have demonstrated the potential for high precision and accuracy; however, they may require large, annotated datasets for training and can be computationally expensive. Recently, convolutional neural networks (CNNs) have emerged as a powerful resource for computer vision tasks, for instance, segmentation, etc. [9]. As compared to the traditional methods, CNNs have shown superior performance, especially in medical image segmentation, by introducing learning features from the data [10].

2. Handcrafted Features-Based Methods

Medical image analysis has made extensive use of handcrafted feature-based techniques, including brain tumor segmentation. These techniques involve the segmentation of images using ML algorithms and the extraction of engineered features that define image qualities [18]. Handcrafted features are divided into three categories: intensity-based, texture-based, and shape-based. Intensity-based features capture the local intensity distribution within the image. These features include statistical measurements such as mean, median, standard deviation, and histogram-based metrics [19]. Intensity-based features are useful in differentiating between normal and abnormal tissue regions due to their distinct intensity profiles. Texture-based features describe the spatial arrangement of intensities and reflect the local patterns in the image. Common texture-based features include the gray-level co-occurrence matrix (GLCM), which captures the frequency of specific pixel value combinations at certain spatial relationships [20]; local binary patterns (LBP), which encode the relationship between a pixel and its neighbors [21]; and Gabor filters, which analyze the frequency and orientation information in images [22]. Texture features can be valuable for characterizing the heterogeneity and complexity of tumor regions. Shape-based features capture the geometric properties of the tumor region, providing information about the tumor’s size, shape, and boundary irregularities. Examples of shape-based features include area, perimeter, compactness, and various moments [23]. These features can help differentiate tumors from surrounding tissues based on their distinct morphological characteristics. ML algorithms, such as random forests (RF), support vector machines (SVM), and k-nearest neighbors (k-NN), are trained for segmentation tasks after handcrafted features are extracted [24]. Despite the success of handcrafted feature-based methods in various medical image segmentation tasks, these methods often require extensive manual tuning and are sensitive to variations in image intensity, limiting their robustness and precision [25]. Additionally, the reliance on manually engineered features can lead to a lack of adaptability to diverse imaging conditions and tumor appearances. Therefore, there is a need for more robust and versatile approaches to tumor segmentation.

3. Convolutional Neural Network-Based Methods

CNN has revolutionized the field of image recognition and segmentation by automatically learning features from the data, making them more robust to variations and alleviating the need for manual feature engineering [17]. CNNs are composed of multiple layers, including convolutional, pooling, and fully connected layers, that use nonlinear transformations to learn hierarchical representations of the input data [26]. This allows CNNs to effectively capture complex patterns and structures within images, leading to improved performance in various image analysis tasks. Several CNN architectures have been proposed in the context of brain tumor segmentation to address the challenges faced by the heterogeneity and complexity of brain tumors. Some of the most prominent architectures include U-Net, V-Net, and DeepMedic ^[27][28][29][27,28,29]. The accurate localization of tumor boundaries is made possible by the symmetric encoder-decoder architecture known as U-Net, which uses skip connections to merge low-level and high-level data. V-Net extends the U-Net architecture to 3D medical images and incorporates a volumetric loss function for improved segmentation performance. DeepMedic employs a multi-scale approach with parallel processing of image patches at different resolutions to capture both global and local contextual information. The study in [30] aimed to accurately segment brain tumors from MRI scans using a 3D nnU-Net model enhanced with domain knowledge from a senior radiologist. The approach improved the model’s performance and achieved high Dice scores for the validation and test sets. The approach was validated on hold-out testing data, including pediatric and sub-Saharan African patient populations, demonstrating high generalization capabilities. These CNN architectures have demonstrated superior performance in brain tumor segmentation compared to traditional methods by learning context-aware features that capture both local and global information [31]. Additionally, CNN-based methods are more robust to intensity variations and can adapt to diverse imaging conditions and tumor appearances, making them a promising approach for this task. Despite the success, CNNs usually need large, annotated datasets for training, which can be challenging to obtain in the medical domain due to the limited availability of expert annotations and the time-consuming nature of manual segmentation [32]. Furthermore, CNNs can be computationally expensive, particularly for large 3D medical images, and may lack interpretability due to their black-box nature. To overcome these limitations, researchers have explored various strategies, such as transfer learning, data augmentation, and incorporating domain knowledge through the integration of handcrafted features. These approaches aim to leverage the strengths of both handcrafted feature-based methods and CNNs to improve the robustness, precision, and interpretability of brain tumor segmentation techniques.

4. Hybrid Approaches in Medical Imaging

Hybrid approaches aim to combine the strengths of handcrafted features and DL techniques to increase the performance of medical image segmentation tasks, taking advantage of domain knowledge and automated feature learning [33]. Several hybrid approaches have been proposed for various medical imaging applications, including lung nodule detection, breast cancer segmentation, and retinal vessel segmentation ^[34][35][36][34,35,36]. These hybrid approaches often involve integrating handcrafted features at different levels of the CNN architecture, such as input channels, feature maps, or decision levels [33]. Several strategies have been proposed for incorporating handcrafted features into DL models. One approach is to concatenate handcrafted features with deep features before the classification layer, which allows the model to leverage both feature types during the decision-making process [36]. Another approach involves injecting handcrafted features into intermediate layers of the CNN, enabling the network to learn more complex, higher-level representations that integrate domain knowledge [37]. Multi-stream architectures, which process handcrafted and deep features in parallel, have also been proposed to encourage complementary learning and robust feature representations [38]. These hybrid approaches have demonstrated an improved performance compared to individually handcrafted features or CNN-based methods in various medical imaging tasks. By combining the advantages of both techniques, hybrid models can capitalize on the domain knowledge provided by handcrafted features while benefiting from the automatic feature learning capabilities of CNNs. Table 1 summarizes a comparison of brain tumor segmentation techniques, including handcrafted feature methods, CNN-based, and hybrid approaches. The evaluation of the relevant literature emphasizes the limitations of handcrafted feature-based methods and the potential of CNN-based methods for tumor segmentation. However, the integration of handcrafted features and CNNs has not been thoroughly investigated for brain tumor segmentation. A hybrid method that combines the strengths of each approach could lead to improved performance in this task, offering a promising avenue for future research and development in the field of medical image analysis.

5. Data Augmentation Techniques

The literature includes an extensive collection of data augmentation in brain MRI. Different techniques were applied to the MRI such as translation, noise addition, rotation, and shearing to increase the size of the dataset as well the performance of tumor segmentation. Khan et al. [39] applied the noise addition to and shearing methods to increase the size of the dataset and improved the accuracy of the classification and tumor segmentation. Similarly, Dufumier et al. [40] applied rotation, random cropping, noise addition, translation, and blurring to increase the dataset size and performance in the prediction of age, and sex classification. Different studies used elastic deformation, rotation, and scaling to improve tumor segmentation and accuracy at the same time [41]. These techniques are common due to their simplicity and performance. In addition to these techniques, the researchers also generated synthetic images to perform a specific task. The most common method of image generation is the mix-up, where the patches from two random images are combined to generate the new image. In all these applications, the researchers used different datasets and different numbers of images. Similarly, everyone used a different network architecture. Thus, every researcher presented the results performance based on their selected techniques. AIn this article, after careful evaluation of the literature, the common techniques are used. These techniques are presented in Table 2. Furthermore, Nelapa et al. [42] provided a comprehensive survey of the data augmentation that can be used for further details.