Enhancing Brain Tumor Segmentation using Berkeley Wavelet Transformation and Improved SVM

1.1. The Significance of Brain Tumor Segmentation in Medical Imaging

Since brain tumor segmentation is essential to the accurate diagnosis and treatment of neurological disorders, it is a highly significant field in medical imaging. With the increasing incidence of brain tumors worldwide, there is a requirement for solid segmentation strategies to improve medical diagnosis and treatment recommendations with increasing precision [1-3]. For the planning of treatment, monitoring of progression, and response-to-treatment assessment, a clinician must be able to delimit the boundaries of a tumor from the surrounding healthy tissues. Brain tumor segmentation represents one of the prerequisites for radiologists and oncologists during medical imaging, where magnetic resonance imaging prevails. Accurate segmentation recovers spatial and morphological information regarding the complete position of the tumor, which further assists medical professionals in describing the form, dimensions, and placement of cancer within the complex structure of the human brain. Whatever the best course of treatment may be, chemotherapy, radiation therapy, surgery, or a combination of these, this information is crucial [4, 5].

Furthermore, the advancement of personalized medicine depends on precise segmentation. No less important is that better and more accurate segmentation allows for more effective and patient specific treatment regimes, individualized to the characteristics of each patient's tumor. This translates easily into two ways in which patient outcomes are improved: all around better quality of life for the treated individual and fewer unnecessary interventions. Researching the natural history of brain tumors and evaluating the efficacy of new treatments in light of studies and clinical trials also require accurate segmentation. Researchers employ robust segmentation techniques to monitor the evolution of tumor morphology and size over time. These findings offer crucial information about how the disease progressed and how well the treatment worked [6-8].

1.2. Motivation and Research Objectives

Techniques such as Berkeley Wavelet Transformations and Support Vector Machines (SVM) were developed to enhance the process of medical image analysis used for brain tumor segmentation with more accurate effectiveness and better results [9, 10]. These techniques could help; however, quite often, the traditional method fails to detect the subtlety among all the major types of brain tumors. The Berkeley wavelet transform offers a persuasive solution in facilitating the extraction of coarse and fine-grain image features and allows multi-level analysis. In the case of medical imaging, it becomes helpful in situations where primitive differences like the tissue could help to recognize the presence of pathological conditions [11, 12].

The optimization target is in line with refining the performance of the classifier algorithm SVM, a robust model applied in a broad sense in the domain of medical image analysis. Though generally highly respected for their capability in binary classification, the problem related to brain tumor segmentation should be taken as more complicated and handled on a more sophisticated and nuanced level. Improving the capability of SVM to detect complex and small changes, particularly in tumor region delineation, must be prioritized to make the process more precise and reliable. High sensitivity, noise, and unpredictability are essential elements of medical images, and the SVM must be tailored to meet their specific requirements [13, 14].

This work's main goal is to combine the benefits of Berkeley Wavelet Transformation with an improved SVM to optimize the brain tumor segmentation process. These cutting-edge techniques aim to enhance diagnosis and treatment planning in the long run by more precisely and consistently segmenting brain tumors.

3.1. Data Collection

The dataset used in this study was compiled from a wide range of medical imaging tests, including brain CT and MRI scans. The collection includes both brain tumor-free and tumor-filled images, among many other kinds of images. The suggested segmentation model must be trained and evaluated using this sizable dataset under various pathological circumstances and anatomical variations. The sample brain tumor images are demonstrated in Fig. (1).

Thoroughly splitting the dataset into training and testing sets made it possible to evaluate the model in great detail. Seventy percent of the 1713 images in the dataset were set aside for training. Based on this large portion of the dataset, the model is able to identify complex patterns and attributes associated with both tumor and non-tumor classes. Three-quarters of the dataset, or 735 images, were set aside specifically for the trained model's testing. This section enhances the reliability of the results and simulates real-world scenarios to offer an unbiased evaluation of the model's generalization capabilities on anonymous data.

3.2. Data Source and Ethical Considerations

The medical imaging data used in this study were obtained from an open-source repository on Kaggle, a widely recognized platform for hosting datasets. The dataset, titled “The BraTS Dataset,” includes anonymized MRI and CT images specifically curated for research purposes. All data are publicly available and were accessed under the terms and conditions outlined by the repository, ensuring compliance with ethical standards. Since the dataset is open-source and does not include personally identifiable information, ethical clearance was not required for this study. The use of this dataset aligns with ethical research practices and supports reproducibility in scientific investigations.

3.3. Dataset Size and Mitigation of Bias

The dataset used in this study, obtained from an open-source Kaggle repository, includes a limited number of MRI and CT images. Since the dataset size is relatively small, we employed several measures to mitigate potential biases and enhance the robustness of our analysis:

(1) Data Augmentation: To increase the effective size of the dataset and improve its diversity, we applied data augmentation techniques such as rotation, flipping, scaling, and random cropping. These methods artificially expanded the training data, allowing the model to generalize better to unseen data and reducing the risk of overfitting.

(2) Cross-Validation: A k-fold cross-validation approach (with k=5) was implemented to ensure robust model evaluation across multiple splits of the data. This technique minimizes the influence of biases that may arise from a single training-test split and provides a more comprehensive assessment of model performance.

(3) Generalizability and Benchmarking: The dataset used is a standard benchmark dataset frequently utilized in similar research studies. While the sample size is limited, its open-source nature enables reproducibility and comparability with related work in the field of medical imaging and machine learning.

(4) Future Considerations: We acknowledge the limitations posed by the small dataset size and its potential impact on the generalizability of the results. Future studies will incorporate larger and more diverse datasets, including multi-institutional and real-world data, to further validate and refine the proposed models.

By implementing these strategies, we aimed to mitigate potential dataset biases and ensure the reliability of the results presented in this study. These efforts have been incorporated into the revised manuscript to enhance its clarity and rigor.

3.4. Pre-processing Steps

The consistency and quality of the medical imaging dataset used to segment brain tumors were improved by meticulous pre-processing. The latter is an important step that deals with intrinsic variations and artifacts in the raw CT and MRI images to attain better performance of the segmentation model. Normalization is also a simple pre-processing step, and the first pre-processing step equates the pixel intensities in every image. This normalization process kept pixel values within a fixed range for all pixels, thus minimizing the effect of intensity variation caused by changes in imaging hardware or imaging protocols. In contrast, pixel value normalization made the model more robust and not very sensitive to the changes in picture acquisition parameters.

All images were after that scaled to make sure spatial resolutions were consistent post-normalization. With the last step in place, the model would be compatible with inputs of constant size, thus ensuring a homogeneous dataset. Resizing allowed images of diversity in the training pipeline, and in the next model training, it increased the computational efficiency. Furthermore, applying noise reduction techniques increases the images' signal-to-noise ratio. Applying methods like median filtering or Gaussian blurring helped suppress extraneous details and promote a cleaner representation of anatomical structures because medical imaging data is prone to noise artifacts. In addition to improving the images' visual clarity, this noise reduction allowed for more accurate feature extraction in the later stages of the segmentation model.

By using intensity normalization techniques specific to medical imaging, potential variations in imaging modalities and intensities were taken into account. For the model to correctly identify relevant features, all pixel values, regardless of the intensity characteristics of the original image, had to be calibrated to a common scale.

Table 1 provides a brief overview of all the necessary pre-processing steps, emphasizing how each one improves, cleans, and standardizes the medical imaging dataset in order to get it ready for brain tumor segmentation.

3.5. Berkeley Wavelet Transformation (BWT)

Berkeley Wavelet Transformation (BWT) is utilized in this work to improve the features used in brain tumor segmentation. The powerful Berkeley Wavelet Transformation image analysis tool allows for multi-resolution decomposition and the extraction of both local and global features from an image. Brain tumors can be segmented using this technique, which also provides a more thorough analysis of the tumor boundaries by identifying minute details and patterns in the medical images.

The Berkeley Wavelet Transform uses a set of variable wavelet filters in size and in more than one direction. Mathematically, they are applied to the original image to comprehensively cover the image domain in the integral domain. The overall decomposition at each level uses wavelet decomposition iteratively to split the image into approximation and detail coefficients. In this process, the high-frequency extraction is performed with the detail coefficients, while the low-frequency information is estimated from the approximation coefficients. For now, that model can recognize the very detailed minute features in the image and the global structures by multi-resolution analysis.

The improved features obtained from the Berkeley Wavelet Transformation help to provide a more discriminative representation of tumor boundaries in the context of brain tumor segmentation. The various features used in this research are shown in Table 2. The extraction of texture information, subtle intensity variations, and spatial relationships within an image are all made possible by multi-resolution analysis. These features are required to precisely define tumor regions. The goal of the research is to maximize the segmentation model's capacity to distinguish between areas with and without tumors by incorporating these improved features into the model's later phases. In the end, this will raise the overall precision and accuracy of the segmentation results.

3.6. Step-by-Step Description of Processes

• Data Collection: We gathered MRI scans from publicly available medical imaging datasets, focusing on both healthy and tumor-affected brain images. (Details discussed in section 3.1).

• Data Normalization: Each image was normalized to ensure uniform intensity distribution across the dataset. Normalization was performed by scaling pixel values to a range of 0 to 1, which helps in reducing the computational complexity and improving model convergence.

• Noise Reduction: To minimize the impact of noise, we applied Gaussian filtering with a kernel size of 3×3. This step enhances the clarity of the images by smoothing out the variations that do not contribute to the detection of tumors.

3.7. Various Machine Learning Models used in This Research

It is strategically possible to successfully and accurately forecast brain tumors with the aid of a gigantic ensemble of machine-learning models. The conventional but enhanced SVM provides a strong foundation for effective classification through its discriminative power.However, in some research, ANN structures are also used, which are strong in identifying complex relationships and patterns. Recurrent neural networks are the best techniques for handling the temporal characteristics and sequential dependencies of medical imaging data. Furthermore, the model uses Convolutional Neural Networks (CNNs) to automatically extract hierarchical features from image data, capturing spatial relationships needed for more intricate predictions [25, 26].

3.8. Significance of Brain Tumor Detection and Rationale for Model Selection

3.9. Computational Demands and Model Interpretability

• Computational Demands

The computational requirements of the machine learning models implemented in this study vary depending on their architecture and complexity. Key observations regarding the computational demands are as follows:

Enhanced SVM and ANN: These models demonstrated relatively low computational demands, making them suitable for deployment in resource-constrained environments. The Enhanced SVM's efficiency arises from its smaller parameter space and focuses on kernel-based feature separation, while ANN's computational simplicity stems from its straightforward architecture.

CNN and RNN: CNN and RNN models require significantly more computational resources due to their complex architectures and the need to process high-dimensional image data. CNNs, with their multiple convolutional and pooling layers, required higher GPU memory for efficient training, while RNNs, particularly those utilizing LSTMs, demanded additional computational power to process sequential dependencies.

To provide a clear perspective, the revised manuscript includes details on the hardware specifications used (e.g., GPU, CPU, memory) and the approximate training and inference times for each model. This information allows readers to assess the feasibility of deploying these models in different computational environments, including clinical settings.

• Model Interpretability

Interpretability is critical for building trust in machine learning models, particularly in medical applications where the outcomes influence clinical decisions. In this study, model interpretability was addressed as follows:

Visualization Techniques: For CNN, saliency maps and Grad-CAM (Gradient-weighted Class Activation Mapping) were applied to visualize the regions of interest in medical images that influenced the model's predictions. These techniques provide clinicians with insights into how the model identifies tumor regions, increasing transparency and trust.

Feature Importance Analysis: For Enhanced SVM, we analyzed the most significant features contributing to classification by examining the weights and kernel outputs. This helps in understanding which features of the tumor (e.g., shape, intensity, texture) were most influential in the decision-making process.

Temporal Analysis in RNN: RNN models were analyzed using sequential visualization techniques to highlight the temporal dependencies that influenced the classification decisions. This approach provides an understanding of how sequential slices of imaging data contribute to predictions.

4.1. Training Data and Model Training Process

4.2. Results and Discussion

Additionally, confusion matrices were used to illustrate the true positive, false positive, true negative, and false negative rates for each model, providing a detailed understanding of their classification performance. Each machine learning model's prediction performance was meticulously assessed after a lengthy multi-feature training phase. Fig. (2) demonstrates the accuracy of the proposed machine-learning model.

The enhanced Support Vector Machine (SVM) won with a remarkable accuracy of 97.6%. The SVM model's enhanced capacity to identify intricate patterns and relationships across a large range of feature spaces makes it a desirable substitute for brain tumor prediction. Second place went to the Convolutional Neural Network (CNN), which achieved an accuracy of 95.76%. CNN's ability to automatically extract hierarchical features from medical images was crucial to its accurate prediction-making. This model has shown such successful diagnostic ability in the diagnosis of brain tumors due to its capability of identifying tiny patterns within the data and finding spatial correlations.

The RNN, while diagnosing, showed relatively high predictive power at a percentage of 92.3. Since the RNN can model data sequentially, it was able to identify temporal dependencies in the medical imaging data. This proved how well-suited the RNN was for sequential data and how it could be used in the dynamic domain to accurately predict brain tumors. Meanwhile, the Artificial Neural Network's (ANN) accuracy stood at 88.77%. Even though the ANN's overall performance was only slightly worse than other models, its capacity to identify complex patterns in the feature space significantly improved its predictive performance. The ability of the ANN to find non-linear relationships and dependencies in the data is still important in the broader context of brain tumor prediction. In summary, the enhanced SVM model outperformed other machine learning models as the most accurate predictor in the study.

Table 3 provides a detailed, complete view of how the performance of the machine learning models predicts the identification of brain tumors in terms of accuracies, precisions, recalls, and F1 scores. The top classifier is the enhanced Support Vector Machine, with an accurate performance of 98.1%, a precision of 96.9%, a recall of 97.5%, and an F1 score of 97.6%. This model showed significant levels of accuracy in the identification of true-positive results, a reduction in false positives, and precise encapsulation of the overall diagnostic performance. The CNN demonstrated the automated hierarchical feature extraction of medical images with scores equal to 96.2%, 95.76%, 96.5%, and 95.8% in the recall, accuracy, precision, and the F1 score, respectively. The predictive analysis indicates that the predictive accuracy, precision, recall, and F1 score of the RNN are at 92.80%, 91.50%, 92.10%, and 92.30%, respectively. This is due to its modeling approach in sequence; therefore, it does well in capturing dependencies that most likely exist in data. With impressive precision, recall, F1 score, and accuracy values of 89.2%, 88.1%, 88.6%, and 88.77%, respectively, the Artificial Neural Network (ANN) demonstrates its ability to recognize intricate patterns in the feature space.

Fig. (3) demonstrates how well each machine learning model performs in terms of precise classification when using confusion matrices to identify brain tumors. In 489 cases of true positive predictions and 941 cases of true negative predictions, the enhanced Support Vector Machine (SVM) has proven its capacity to accurately identify tumors and non-tumors. However, eleven actual tumors are mistakenly classified as non-tumors (false negatives), and nine non-tumors are mistakenly labeled as tumors (false positives).

Similarly, the Convolutional Neural Network (CNN) achieves 482 true positives and 937 true negatives but misclassifies 18 tumors and 13 non-tumors. Recurrent Neural Network (RNN) produces 468 true positives and 929 true negatives by incorrectly classifying 21 non-tumors and 32 tumors. The Artificial Neural Network (ANN) achieves 454 true positives and 919 true negatives, with 46 tumors and 31 non-tumors misclassified. These matrices offer a comprehensive understanding of the model's performance by highlighting the exact and accurate predictions. Additionally, they provide helpful data that can be used to enhance and optimize each model in the intricate process of identifying brain tumors.

4.3. Validation and Generalization

We acknowledge that the findings of this study have not been validated on independent datasets, which is a limitation in demonstrating the generalizability of the proposed models across diverse populations and imaging protocols. The dataset used in this research was obtained from an open-source repository on Kaggle and served as a foundational benchmark for developing and evaluating the machine learning models. However, reliance on a single dataset restricts the scope of generalization.

To address this limitation and enhance the robustness of the findings, we propose the following steps for future work:

(1) Validation of Independent Datasets: Future research will focus on validating the proposed models on independent datasets collected from different institutions and regions. This will allow us to evaluate the performance and adaptability of the models across diverse populations and imaging modalities.

(2) Incorporation of Multi-Institutional Data: Expanding the dataset to include multi-institutional data with varied patient demographics, imaging techniques, and tumor characteristics will help reduce potential biases and improve the generalizability of the findings.

(3) Transfer Learning and Domain Adaptation: To adapt the models for different datasets, we plan to utilize transfer learning techniques. Fine-tuning pre-trained models on new datasets will ensure better performance and applicability to broader use cases.

(4) Real-World Validation: Collaborative efforts with healthcare institutions will be pursued to test the models on real-world clinical datasets, ensuring the models' practical utility in clinical settings.

These future directions have been outlined to ensure that the limitations identified in this study are addressed in subsequent research. By validating the findings on independent and diverse datasets, we aim to establish the robustness and clinical relevance of the proposed models in the field of medical imaging and brain tumor detection.

4.4. Need for Clinical Validation

While the machine learning models developed in this study demonstrate promising results in the detection of brain tumors using publicly available datasets, we acknowledge that clinical validation is essential to ensure their practical applicability in real-world healthcare settings. Clinical validation will provide a robust assessment of the models' performance under actual clinical conditions and further establish their reliability for diagnostic use.

To address this critical aspect, the following steps are proposed for future work:

(1) Collaborations with Healthcare Institutions: We aim to collaborate with hospitals and medical research centers to test the trained models on clinical datasets. These datasets will include real-world patient imaging data, encompassing diverse demographics and imaging protocols, to evaluate the models' performance and generalizability.

(2) Prospective Clinical Studies: Future research will involve conducting prospective studies where the trained system is integrated into the clinical workflow. This will allow us to assess its ability to support diagnostic decisions in real-time under the supervision of medical professionals.

(3) Integration of Clinician Feedback: During the clinical validation process, feedback from radiologists and oncologists will be gathered to refine the system. This iterative process will help ensure that the model's predictions align with clinical expectations and provide actionable insights for medical practitioners.

(4) Adherence to Regulatory Standards: The models will be validated following regulatory guidelines, such as those provided by the FDA or CE, to ensure compliance with the standards required for clinical deployment. This will also involve ensuring data privacy, ethical use, and reliability of the system.

By undertaking these steps, we aim to provide comprehensive clinical validation of the proposed system, ensuring its readiness for deployment in medical diagnostics. These efforts have been outlined as part of the future direction of this study, reinforcing our commitment to bridging the gap between research and clinical practice.