V3I7P74

Enhancing Brain Tumor Classification in MRI Using Multi-Backbone Deep Features and Radiomic Fusion

Safaa K. Hussaine¹*

Abstract

Accurate detection of brain tumours from magnetic resonance imaging (MRI) is important for diagnosis, treatment planning and prognosis. Here, we propose a hybrid machine learning/deep learning framework that combines handcrafted radiomic descriptors with deep features extracted using pretrained convolutional neural network backbones, including DenseNet121, Xception and ResNet50. Radiomic and deep features were fused in a unified representation and classified by support vector machine (SVM) and multi-layer perceptron (MLP) classifiers. As an end-to-end baseline we trained a lightweight convolutional neural network on raw MRI images. The framework was evaluated on a four-class brain MRI dataset, which includes glioma, meningioma, pituitary tumour and no-tumor images. The hybrid SVM model reached the best classification accuracy of 98.63% as opposed to 96.95% of hybrid MLP and 83.07% of the CNN baseline. The SVM also demonstrated nearly perfect class specific receiver operating characteristic area-under-the-curve values. These results suggest that the combination of radiomic descriptors with pre-trained deep representations increases class discrimination and classification performance with respect to the selected end-to-end CNN baseline. Thus, the proposed framework offers a promising direction for automated multi-class brain tumour classification from MRI images.

Keywords:

Brain tumor classification; magnetic resonance imaging; radiomics; DenseNet121; Xception; ResNet50; support vector machine; multi-layer perceptron; feature fusion.