A New Framework for Brain Tumor Feature Extraction and Classification Using Localized Global Feature Patches

Texture-based feature extraction is crucial in brain tumor feature extraction and classification. The texture-based feature extraction provides valuable information on the textural difference between the tumorous and non-tumorous region and its spatial arrangement of pixel intensities. In this proposed framework, machine learning models using Gray-Level Co-occurrence Matrices (GLCM) and Gabor filter on localized global feature patches have been developed to classify different types of brain tumors. The best results were obtained using GLCM features: energy, contrast, correlation, and homogeneity. The concatenation of the features of localized $35 \times 35$ patches reduces the computational requirements and complexity. The proposed framework provides robust methods for characterizing various texture-based features using localization. Figshare, a publicly available dataset of 3064 images, has been used in this work for three classes: Meningioma, Glioma, and Pituitary. The no-tumor dataset has been obtained by publicly available Br35H, making 4010 images. The dataset is divided into 70:30, 80:20, and 90:10. The best result is established with the split of 70: 30 in comparison to 80: 20 and 90: 10. The features are selected using the Atrous Convolution Autoencoder (A-CAE). The dimensions have been reduced by 42%. The selected features are then provided as input to the Support Vector Classifier (SVC), Random Forest (RF), K-Nearest Neighbour (KNN), and Extreme Gradient Boost(XGBoost). Out of which, the XGBoost classifier on hyperparameter tuning has obtained the highest accuracy of 91.83%.