A Hybrid Clustering Approach: Segmentation and Classification of Brain Tumour Utilizing SVM and CNN Methods

The important and complicated structure of the human anatomy is the brain. Billions of cells work cohesively to perform a task, making its function complex. When there is a blunt dissection of the brain cells, leading to an anomalous collection of cells inside or around the brain, it leads to brain tumour. This group of anomalous cells bothers the regular activities of the brain and abolishes the healthy tissues. The two categories of brain tumours are categorized into two types benign or low grade (Grade I & II) and malignant or high grade (Grade III & IV). Benign tumours are non-cancerous, which grows gradually and does not spread in the tissues because it is less destructive [1].

To detect the abnormalities accurately, MRI is widely recognized as a superior imaging tool that captures precise and detailed organs from inside of the body [2, 3]. The different types of MRI, like the high field MRI, which helps us capture the good images, and the low field MRI to capture low field category MRI images. MRI imaging technique helps physicians to visualize the hairline cracks and tear in muscles, certain soft tissue, and ligaments that occur during injuries. MRI is a progressive medical imaging technique that provides detailed information about human soft-tissue anatomy. In the process of automatically segmenting brain tumours, abnormal tissues can be distinguished from normal structures such as white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF) [4, 5]. When comparing MRI with other methods, it uses appropriate Radiofrequency and gradient pulses with suitable relaxation timing to produce different image components. The techniques are quite relevant to produce constant high-quality images.

The life expectancy of the tumour patients is only two years in the case of LGG and HGG glioma tumours. Several imaging techniques were utilized, such as MRI, CT, Positron emission tomography (PET), and single-photon emission tomography (SPECT). To diagnose and treat glioma, the MRI is the more proper imaging practice. Because the MRI imaging technique provides high resolution while capturing soft tissues, multi-parameter computation, arbitrary direction adjustments, noninvasive imaging, etc. Various Imaging types provide exhaustive information on images and designate the characteristics of tumour [6]. It is a challenging problem to partition the tumour portion accurately due to the shape, location, appearance, and dimensions of gliomas that can change from patient to patient. The second region is gliomas surrounded by invading tissues, making the boundary blurred. The segmentation is becoming complex due to imaging distortion due to imaging devices and the capturing methods.

The brain tumour is in different sizes and shapes; it is hard to perform segmentation because MRI images have information at multiple scales due to their broader structure. Hence, to capture the information at various levels, decomposition of the images into different frequency bands was done using DWT. MRI images are a huge dimensional dataset; hence, it takes more values to represent the voxel in the spatial dimension. Dimensionality reduction using DWT-based PCA reduces the computation time further and helps to avoid overfitting. It is more effective in denoising MRI images, which enhances the robustness of feature extraction. The proposed model uses multiscale feature extraction, dimensionality reduction, noise robustness, and hybrid clustering to capture subtle variations in tissue characterization, allowing for more accurate segmentation results.

The proposed methodology uses a CNN and SVM-based architecture to identify the tumour from an MRI brain image. The implemented technique comprises the following steps for brain tumour classification. The suggested work is depicted in the overview provided in Figure 1.

Step 1: Data set collection and preprocessing.

Step 2: Segmentation of tumour is achieved by hybrid clustering.

Step 3: Utilize DWT and PCA for feature extraction to reduce dimensionality.

Step 4: CNN and SVM based classification and analyse the performance of the two classifiers.

101.jpg

Figure 1. Process flow diagram

3.1 Data set and preprocessing

The 2020 BRATS training dataset contains 399 multi-modality MRI scans, including 293 from glioblastoma (GBM/HGG) and 76 from lower grade glioma (LGG), along with their respective ground truth segmentations for evaluation.

The image size is 512×512 pixels with a dimension of 0.49×0.49 mm². Data augmentation and random transformations are used to reduce overfitting of the neural network in the training phase. 80% used for training, while the rest 20% for testing. The data augmentation technique is employed to provide robustness, preserve spatial features from MRI brain images, and increase the precision and reliability of the detection process. Rotating operations is a common data augmentation technique used in brain tumour segmentation. Rotating operations are applied to the input images during each epoch of training. It helps the CNN model learn features that are relevant to such transformations. Additional training samples are generated using random flipping along with the original dataset. The CNN and SVM models are trained using the augmented dataset, which helps the models become more robust and reduce overfitting. Additionally, it helps to increase the generalization capabilities of the CNN and SVM models. Dropout layers are used before the fully connected layer to reduce overfitting.

With respect to the class imbalance present in the BRATS dataset, which has a significantly higher number of HGG cases than LGG cases (293 vs. 76), we covered this gap with strong data augmentation, like rotating and flipping to increase the amount of LGG samples in the training set. With the application of these augmentations in dynamic epochs, the model was able to further improve generalization, mitigating overfitting. While we did not actively apply weighting loss functions or oversampling, the augmenting approach provided passive albeit effectively balanced, inputs. For class imbalance with the aim of reinforcing segmentation performance in both types of tumors, further methods like class-weighted loss and synthetic over-sampling (e.g., SMOTE) will be explored in later work.

3.2 Image binarization

Binarization, a fundamental image processing step, involves converting a grayscale image into a binary image. The two possible values for each pixel, often zero and one as in the equation, are called a binary image (I) (I :(x, y) à (0,1)). Binarization refers to the fundamental process within an image that helps separate the foreground object from the background, as shown in Figure 2.

102.jpg

Figure 2. Convolutional neural network

The proposed research work uses Otsu’s based thresholding method. The main goal of this method is to find the optimal threshold value.

This approach involves grouping pixels into two corresponding classes, C1 and C2, based on a bimodal histogram, aiming to select thresholds that minimize intraclass variance using a weighted equation for each cluster.

$\sigma^2 w(t)=\mathrm{q}_1(\mathrm{t}) \sigma_1^2(\mathrm{t})+\mathrm{q}_2(\mathrm{t}) \sigma_2^2(t)$           (1)

where,

The weight q_i  is the probability for each class.

It can be considered as follows

$\begin{gathered}\mathrm{q}_1(\mathrm{t})=\sum_{i=1}^t P_i \\ \mathrm{q}_2(\mathrm{t})=\sum_{i=t+1}^l P_i \\ \mu_1(\mathrm{t})=\sum_{\mathrm{i}=1}^{\mathrm{t}} \frac{{ }^{\mathrm{iP}}(\mathrm{i})}{\mathrm{q}_1(\mathrm{t})} \\ \mu_2(\mathrm{t})=\sum_{\mathrm{i}=\mathrm{t}+1}^{\mathrm{l}} \frac{\mathrm{iP}(\mathrm{i})}{\mathrm{q}_2(\mathrm{t})}\end{gathered}$

The distinct class variance is specified by

$\begin{aligned} \sigma_1^2(\mathrm{t}) & =\sum_{\mathrm{i}=1}^{\mathrm{t}}\left[\mathrm{i}-\mu_1(\mathrm{t})\right]^2 \frac{\mathrm{P}_{\mathrm{i}}}{\mathrm{q}_{1(\mathrm{t})}} \\ \sigma_2^2(\mathrm{t}) & =\sum_{\mathrm{i}=\mathrm{l}+1}^{\mathrm{t}}\left[\mathrm{i}-\mu_2(\mathrm{t})\right]^2 \frac{\mathrm{P}_{\mathrm{i}}}{\mathrm{q}_{2(\mathrm{t})}}\end{aligned}$

After the variance is computed, the procedure is terminated

$\sigma_{\mathrm{b}}^2(\mathrm{t})=\sigma-\sigma_{\mathrm{w}}^2(\mathrm{t})=\mathrm{q}_1(\mathrm{t})-\mathrm{q}_2(\mathrm{t})\left[\mu_1(\mathrm{t})-\mu_2(\mathrm{t})\right]^2$                (2)

The technique is employed to minimize intra-class variance while maximizing between-class variance.

3.3 Image segmentation

The process of dividing the entire image into smaller regions is called segmentation. It is necessary to segment the features in an image since distinct features require focus and clustering falls under simple unsupervised learning. Grouping of pixels with similar intensity is known as clustering, which is done without any training images. Hybrid segmentation refers to the process of integrating two segmentation algorithms.

The purpose of combining various algorithms is to eliminate the shortcomings of two dissimilar methods and to improve the good segmentation results. The result is shown in the Figure 3. The combination of fuzzy C-means with K-means is called the KIFCM technique, which is used for accurate tumour detection from MRI images [14]. By integrating Fuzzy C-Means and K-Means, the KIFCM technique improves tumor detection accuracy in MRI images due to the shortcomings of each method. It increases segmentation precision and decreases computational time. Careful segmentation design considerations are necessary to ensure quality segmentation results.

103.jpg

Figure 3. Segmentation results (a) Input image (b) Otsu’s threshold image (c) Segmentation of tumour

3.4 Feature extraction

The computational complexity of processing large dataset was reduced by feature reduction.

The extracted features, including intensity and texture-based ones such as contrast, energy, and entropy, aid in classifying MRI brain images into dissimilar regions like GM, WM, CSF, and tumour.

3.5 Discrete wavelet transform

The different frequencies of an image were analyzed with the help of a wavelet. DWT is one of the tools for feature extraction. Wavelet helps in localizing frequency information important for classification, with spatial frequency components acquired from LL and HL sub-bands, which describe image texture features. The statistical features are obtained using gray level co-occurrences matrix (GLCM). The statistical features are obtained using GLCM and also gray-level spatial dependence matrix (GLSDM). The proposed research work uses GLCM to extract features like contrast; correlation, energy, homogeneity, entropy, and variance were obtained from LL and HL sub bands of the first four levels of wavelet decomposition.

3.6 Principal component analysis

The main use of PCA is dimensionality reduction. Feature extraction can be performed on both the training MR image and testing MR image. During testing and training phase, feature extraction can be done. During the training phase, the features are removed from each training MR image [15]. The training image 1 can be represented by and the pixel resulting is M*N (M rows and N columns).

The PCA algorithm reduces the dimensionality of feature vectors for both training and test images. It computes the Euclidean distance similarities among the reduced feature vectors of test and training images for classification.

The MRI image [16] is correctly classified if i =j; otherwise, i it is misclassified.

MRI images are multiscale, and they hold information that can be retrieved by DWT in frequency bands as images are decomposed. Because of the high dimensionality of MRI data, computation time, overfitting, and noise, PCA is employed post-DWT to enhance these attributes. Although DWT ensures the frequency components that are non-linear and relevant to the tumor are preserved, which is why PCA isn’t a problem. Then DWT ensures that the most informative features based on variance are retained. Unlike t-SNE or Autoencoders, PCA does not require training, works well with DWT, is more efficient, and is not time-consuming. This leads to the preservation of vital tumor characteristics, which leads to improved segmentation accuracy.

The research work compares MRI brain tumour segmentation using various methods such as Random Forest, Decision Tree, U-Net, and Seg-Net. The performance of segmentation relies on handcrafted features like intensity and texture features. However, SVM struggles with capturing complex spatial relationships and variations in MRI brain images compared to deep learning methods. Random Forest, along with other segmentation processes, provides ensemble learning approaches that build multiple decision trees during training, yet it faces challenges in capturing spatial dependencies effectively. Decision Tree, when employed for MRI brain tumour segmentation with a single tree, struggles to capture complex patterns. Another well-known deep learning architecture is U-Net, which has a symmetric expanding path for accurate localization and a controlling path to capture context.

The skip connections of the U-Net help retain high-resolution features, resulting in superior performance compared to conventional CNNs, albeit achieved with a smaller dataset. Comparatively, CNN-based methods for MRI brain tumour segmentation require larger datasets. However, the SVM method outperforms with fewer datasets. The choice of the method depends on various factors such as computational resources, data availability, and the specific requirements of the segmentation task. The proposed research work used in BRATS 2020 dataset. The SVM model outperforms for smaller datasets with less computational power and provides clear decision boundaries. However, the CNN model requires a larger dataset for training and higher computational resources. The practical deployment aspects are very important. Steps taken to increase performance allowed us to add a discussion on the model interpretability issues (particularly with SVMs) and propose later using XAI approaches, like SHAP, to support trust by clinicians. On the other hand, our system demonstrates low-latency inference on mid-level GPUs, thus, real-time segmentation is achievable. Figure 4 shows a bar chart.

104.jpg

Figure 4. Bar chart

The authors are thankful to the reviewers for their insightful feedback and the invaluable suggestions to improve the initial version of this manuscript. The authors also thank the anonymous assistant for their help with performing the statistical analysis.