An Integrating CNN and Texture Features for Enhanced Brain Tumor Diagnosis from MRI Scans

Abstract

Using T1-weighted contrast-enhanced MRI scans, this paper offers a hybrid model combining deep convolutional features from ResNet50 with manually created texture descriptors for brain tumour classification. Three tumour kinds in the dataset are meningioma, glioma, and pituitary. Steps in preprocessing include converting photos to greyscale, applying median filtering to lower noise, downsizing to a consistent input size, normalising pixel values, and doing data augmentation to increase generalisation. The model's backbone is ResNet50, a pretrained deep convolutional neural network that uses residual learning to retrieve high-level features. At the same time, Gabor filters and the Gray-Level Co-occurrence Matrix (GLCM) are used to extract manually created texture characteristics, which records spatial connections of pixel intensity. Haralick features are generated from GLCM to measure texture patterns, hence complementing deep features. The hybrid model has a dual-branch architecture: one branch handles GLCM-based descriptors, while the other processes ResNet50-derived features. Final classification results from passing these features through fully linked layers with dropout for regularisation and concatenation. By way of contrast, GLCM characteristics alone train conventional machine learning classifiers like Support Vector Machine (SVM) and K-Nearest Neighbours (KNN). To keep balanced class representation, the dataset is divided using an 80:20 stratified train-test split. Model performance is measured and compared using evaluation criteria including accuracy, precision, recall, F1-score, and ROC-AUC. The findings reveal that integrating deep and handmade characteristics improves classification accuracy, hence highlighting the use of hybrid techniques for brain tumour identification from MRI pictures.