1. Introduction

Brain tumors pose a substantial public health challenge worldwide due to their detrimental effects on individual and societal health. These abnormal proliferations within the brain are associated with serious physical health effects, cognitive impairment, and substantial financial burden on families. The brain tumor perception technology, particularly the application of detection and segmentation, has significant potential for mitigating the impact of brain tumors. These techniques, with their roles in early prevention and diagnostic assistance, are critical to improving patient prognosis and informing the most appropriate treatment strategy.

Traditional image detection and segmentation techniques, such as region growing, clustering, and edge detection, have played an important role in brain tumor perception. Further, techniques based on Markov random fields [1, 2] and active shape models [3] have provided reliable methods for identifying and delineating tumor regions in brain scans. However, the complexities and variances inherent in brain tumors present challenges to these traditional techniques.

Machine learning (ML), especially deep learning (DL), have substantially expanded the capabilities of these traditional image processing algorithms. ML models, which include Support Vector Machines (SVM) and Random Forest (RF), showed potential in brain tumor detection, but limitations arose from their dependency on manual feature extraction [4]. In addition, deep neural network algorithms, especially convolutional neural networks (CNNs), introduced an automated framework for feature extraction and learning, thereby enhancing the accuracy of brain tumor detection and segmentation [5, 6]. Subsequent advancements have led to the development of deeper models, such as Visual Geometry Group Network (VGGNet) [7], and Residual Network (ResNet) [8], which further improved the capabilities of these systems. One significant development in DL-based segmentation has been the U-Net architecture, which was widely used in medical image segmentation and has shown impressive results in brain tumor segmentation [9].

The integration of ML and DL in brain tumor perception is having profound effects, with the potential for significantly improved patient outcomes and novel approaches to treatment strategies. Therefore, this paper gives a comprehensive overview and analysis of ML algorithms and their technical progress for brain tumor detection and segmentation. In Section 2, we discuss data acquisition and preprocessing. Sections 3 and 4 delve into ML algorithms for detection and segmentation, respectively. Section 5 explores the integration of these models, and Section 6 provides a conclusion that summarizes our main findings, discusses challenges and limitations, and looks towards future directions.

2. Data acquisition and preprocessing

Accurate and reliable data acquisition and preprocessing are vital in the detection and segmentation of brain tumors. In this section, we will discuss the key considerations and techniques involved in acquiring and preparing the data for analysis. The information presented here is based on real-world studies and research in the field of brain tumor analysis.

2.1. Data acquisition

Magnetic resonance imaging (MRI) is a primary approach for analyzing brain tumors [10]. MRI scans provide detailed information about the brain's structure and offer insights into the presence and characteristics of tumors.

To acquire MRI data, several parameters need to be carefully set to ensure optimal image quality. These parameters include the type of MRI sequence, slice thickness, field of view, repetition time, and echo time [11]. The choice of sequence depends on the specific imaging goals, as different sequences highlight various tissue properties and tumor characteristics.

In addition to MRI, computed tomography (CT) and positron emission tomography (PET) can also provide valuable information for brain tumor analysis. CT scans offer detailed anatomical information, but they have lower soft tissue contrast compared to MRI. PET scans can provide functional information by visualizing metabolic activity in the brain.

2.2. Data preprocessing

Data preprocessing plays a crucial role in image analysis. It involves several steps to enhance image quality, remove artifacts, and normalize the data. The following are common preprocessing techniques applied to brain tumor imaging data:

1) Image Registration: Image registration is to register images in different modes or at different times. It corrects for patient motion during scanning and ensures that corresponding anatomical structures are spatially aligned. Registration techniques involve matching features or optimizing a similarity measure to estimate the transformation between images [12].

2) Noise Reduction: MRI images are often affected by noise, which can degrade image quality and affect subsequent analysis. Various denoising techniques, such as Gaussian smoothing, median filtering, or wavelet-based methods, can reduce noise while preserving important details of the image [13].

3) Intensity Normalization: Intensity normalization is crucial for ensuring consistent image intensities across different scans and subjects. It involves scaling the image intensities to a common range or applying histogram equalization to enhance the contrast [14].

4) Skull Stripping: Skull stripping, or brain extraction, is the process of removing non-brain tissues from the image to isolate the brain region. This step is important to eliminate unwanted structures and artifacts that may interfere with subsequent analysis [15].

5) Bias Field Correction: Bias field correction is necessary to compensate for the gradual spatial intensity variations introduced by the imaging process. These variations can arise from imaging artifacts, scanner inhomogeneities, or acquisition protocols. Correcting the bias field helps to normalize the image intensities and improve the accuracy of subsequent analysis [16].

6) Image Resampling: In some cases, it may be necessary to resample the images to a common resolution or voxel size. Resampling ensures consistent spatial dimensions across different images and facilitates comparison and analysis [17].

7) Data augmentation: Data augmentation technology is aimed at addressing the problem of a dataset that is not rich by adjusting the shape of the image to expand the dataset. Common operations include rotation, scaling, flipping, and elastic deformation [18].

2.3. Quality control and annotation

After preprocessing, it is essential to perform quality control checks to ensure the integrity and reliability of the data. Visual inspection can help identify any remaining artifacts, image distortions, or registration errors that need to be addressed.

For supervised learning tasks such as tumor segmentation, manual annotation of the tumor regions is required. Expert radiologists or clinicians carefully delineate the tumor boundaries on the preprocessed images. The annotations serve as ground truth labels for training and evaluating the segmentation algorithms [19].

2.4. Data splitting and cross-validation

In the verification process of ML algorithm, the dataset is usually divided into training, validation, and testing sets. In order, they are used to train models, adjust hyperparameter, and evaluate performance [20].

In conclusion, data acquisition and preprocessing are crucial steps in brain tumor detection and segmentation. The choice of imaging modality, acquisition parameters, and preprocessing techniques significantly impact the quality and suitability of the data for analysis. Careful attention to these steps ensures accurate and reliable results, laying the foundation for subsequent ML algorithms and analysis pipelines.

3. Brain tumor detection technology

3.1. Classical machine learning techniques

ML has made significant strides in the past few decades, and has been applied in many fields. Brain tumor detection is an important field in which ML is widely used. Initial applications leaned heavily on traditional ML techniques such as Support Vector Machine (SVM) and Random Forest (RF) for classification jobs [21].

SVMs function as binary linear classifiers. They employ a hyperplane in an n-dimensional space to distinguish between classes. In brain tumor identification, SVMs have been used to detect tumors from features derived from MRI scans, such as shapes, textures, and image intensities.

RF, on the other hand, is a collective learning method. It constructs numerous decision trees during training and chooses the class most commonly predicted by the ensemble of trees. In simpler terms, each tree makes a prediction, and the class with the majority votes is selected by the model. RF has proven effective in brain tumor detection by minimizing overfitting and enhancing generalization capabilities [22]. The strength of RF lies in its ability to manage high-dimensional data and maintain an efficient bias-variance balance.

3.2. Classical machine learning techniques

Traditional ML techniques, including SVMs and RF, showed promise in brain tumor detection, but they're not without limitations. They rely heavily on manually crafted feature extraction and grapple with the complexities of high-dimensional medical imaging data. A breakthrough came in the form of DL methods, notably CNNs, which have revolutionized brain tumor detection. CNNs, with their innate ability for hierarchical feature learning, have excelled at identifying intricate patterns in imaging data [23]. By learning multiple representation levels that correspond to different abstraction levels, CNNs can detect tumors with high accuracy. These levels, more commonly referred to as “layers” in DL terminology, represent different abstraction tasks. Each layer learns different features of the data, with the complexity of the features increasing with the layer depth. The first layer might learn simple patterns such as edges in an image, while deeper layers combine these simple patterns to learn more complex ones such as shapes or objects. Thus, each layer corresponds to a different "abstraction level", with deeper layers representing higher levels of abstraction. Specific CNN architectures like AlexNet [23], VGGNet [24], and ResNet [25] have been modified for brain tumor detection tasks, yielding impressive results. These models' power lies in their capacity to learn distinguishing features directly from raw imaging data, negating the need for manual feature extraction and selection, which are prerequisites in traditional ML techniques.

3.3. Performance metrics and comparison

Performance evaluation is an integral part of ML model development. Common evaluation indicators mainly include accuracy, sensitivity, specificity, and area under the Receiver Operating Characteristic (ROC) curve (AUC-ROC). CNN-based models generally outshine traditional ML models on these metrics, boasting higher accuracy and AUC-ROC scores. However, the performance of models can be considerably swayed by the training dataset, underlining the need for cross-validation and robust model evaluation practices.

To sum up, traditional ML techniques have made considerable inroads in brain tumor detection. With the emergence of CNNs, their detection performance has been over the traditional methods. Future research should concentrate on refining these models and exploring how the strengths of both traditional and DL methods can be synergistically combined.

4. Machine learning algorithms for brain tumor segmentation

A single brain MRI scan provides a myriad of slices across a 3D perspective. Therefore, manual segmentation of brain tumor from magnetic resonance (MR) images is a challenging and time-consuming task. Therefore, manual segmentation of brain tumor is a very difficult task in MR images. In addition, the manual segmentation process may introduce errors that impact the diagnosis and treatment decisions for the patient. In order to address the above challenges, many scholars have conducted in-depth research and proposed many methods. These automated brain tumor segmentation methods can help to enhance precision, reduce time consumption, and minimize human error.

This section provides a comprehensive review of research papers on automated brain tumor segmentation techniques. This assessment will primarily focus on four key categories: threshold-based and edge detection methods; region-based methods; statistical-based methods; and ML and DL methods. Each segmentation technique will be carefully reviewed in terms of the experimental datasets used, the particular segmentation algorithm implemented, and the performance metrics reported in the respective studies.

4.1. Methods based on feature extraction

ML has made significant strides in the past few decades, and has been applied in many fields. Brain tumor detection is an important field in which ML is widely used. Initial applications leaned heavily on traditional ML techniques such as SVM and RF for classification jobs [26].

4.2. Methods based on traditional image segmentation

Region-based methods, such as region growing and watershed segmentation, segment images by grouping similar pixels into homogeneous regions [27]. These methods perform well with high-contrast images but struggle with noise and low-contrast scenarios. For instance, region-growing methods can cause over-segmentation in cases of intensity variations within the tumor [28].

4.3. Model based on statistics

Statistical-based methods, including Markov Random Field (MRF) and Active Shape Model (ASM), utilize statistical data for segmentation. MRFs model the spatial dependencies of image pixels, offering robustness against noise [29]. Conversely, ASM uses statistical shape analysis for segmentation, which can be highly effective but is heavily reliant on the quality of the training dataset [30].

4.4. Methods based on machine learning and deep learning

ML and DL methods have gained substantial attention due to their ability to handle complex patterns and large data volumes. Algorithms like SVM and RF have been utilized for brain tumor segmentation [31]. However, these traditional ML techniques require manual feature extraction, which can be laborious and error-prone.

DL, particularly CNNs, has shown excellent performance in brain tumor segmentation. CNNs can autonomously extract necessary features from images and perform learning [32]. The U-Net architecture, for instance, has been a popular choice for medical image segmentation due to its ability to localize features at different scales [33].

4.5. Model based on statistics

In summary, the evolution from traditional segmentation methods to ML and DL techniques represents a significant advancement in brain tumor segmentation. These newer methods can handle complex patterns and large volumes of data, increasing both the efficiency and accuracy of brain tumor diagnosis. Despite the significant progress in the field, challenges such as the necessity of large annotated data sets for training and model interpretability remain. Future research in this area should focus on tackling these issues to make these powerful tools more widely accessible and applicable in clinical practice.

5. Integration of detection and segmentation models

In recent years, there have been significant advancements in ML algorithms for brain tumor detection and segmentation. Traditionally, these tasks were treated as separate processes, with detection algorithms identifying the presence of tumors and segmentation algorithms outlining the tumor boundaries. However, researchers have recognized the benefits of integrating these two tasks into a single framework, leading to more accurate and efficient tumor analysis. In this section, we discuss the advancements in integrating detection and segmentation models for brain tumor analysis.

5.1. End-to-end deep learning models

End-to-end DL models have gained immense popularity due to their ability to jointly perform tumor detection and segmentation. These models leverage CNNs to automatically learn discriminative features and spatial relationships from the input data. By integrating detection and segmentation into a unified framework, these models eliminate the need for separate pipelines and simplify the overall analysis process.

One notable approach in this category is the U-Net architecture [34]. The U-Net model consists of an encoder-decoder structure, where the encoder captures contextual information from the input image, and the decoder generates pixel-wise segmentation masks. The U-Net architecture has shown promising results in various brain tumor detection and segmentation tasks [35, 36].

5.2. Models combining multi-modal image

Brain tumor analysis often requires the integration of multiple imaging modalities, such as magnetic resonance imaging (MRI) scans with different contrasts. Combining multiple modal images for collaborative analysis can effectively compensate for the shortcomings generated by images from different single pathways. More comprehensive image information can effectively help doctors make more accurate judgments.

One approach to multimodal modeling is to fuse the feature representations learned from each modality. For instance, Kamnitsas et al. proposed a 3D CNN architecture that combines the features extracted from T1-weighted, T1-weighted post-contrast, T2-weighted, and Fluid-attenuated inversion recovery (FLAIR) images for brain tumor segmentation [37]. By incorporating multiple modalities, these models can capture a more comprehensive representation of the tumor characteristics.

5.3. Transfer learning and domain adaptation

This technology have shown outstanding development prospects in improving the performance of brain tumor detection and segmentation. These techniques leverage knowledge learned from domain source (e.g., a large annotated dataset) and transfer it to domain target (e.g., a new dataset with limited annotations). By doing so, these models can effectively learn from limited labeled data in the target domain and generalize well to new unseen cases.

For example, Havaei et al. proposed a method that uses a pre-trained CNN on a large-scale dataset to initialize the weights of the network for brain tumor segmentation [38]. They then fine-tuned the network using a smaller annotated dataset from a different domain, achieving competitive results with significantly reduced annotation efforts.

5.4. Brief summary

The integration of detection and segmentation models in brain tumor analysis has shown great promise in improving accuracy and efficiency. End-to-end DL models provide a unified framework for simultaneous detection and segmentation, while multimodal models combine different imaging modalities to capture comprehensive tumor characteristics. Transfer learning and domain adaptation techniques further enhance performance by leveraging knowledge from related domains. These advancements pave the way for more accurate and reliable diagnosis and treatment planning for brain tumor patients.

6. Conclusion

Brain tumor detection and segmentation have crucial implications in medical diagnostics, significantly influencing therapeutic strategies and patient outcomes. This comprehensive review focuses on the advancements in ML algorithms in these areas. Traditional image processing techniques, while valuable, have limitations in the face of complex and variable brain tumors. The integration of ML, particularly DL, has greatly enhanced the capabilities of these traditional techniques, with DL models such as CNNs showing impressive results in both detection and segmentation tasks. Despite the significant progress, challenges persist, particularly in data quality and diversity, model interpretability, and generalizability. The future promises exciting developments in this field, and further research is essential to realize the full potential of these advanced ML algorithms in brain tumor diagnosis and treatment.

Authors contribution

All the authors contributed equally and their names were listed in alphabetical order.