Multi‐Modal Deep Feature Extraction and Classifier‐Level Integration for Brain Tumour Classification Using CT and MRI Image

ABSTRACT An organised and labelled multi‐modal deep learning framework for brain tumour detection using MRI and CT image datasets is proposed. Standardisation, noise reduction, and data augmentation were applied to ensure high‐quality input data and reduced modality‐specific artefacts. For feature extraction, AlexNet, GoogleNet, SqueezeNet, and ResNet captured complementary information from CT and MRI images. Classifier‐level combinations were used to compare CNN backbone features to better represent tumour characteristics. The extracted and fused features were evaluated using SVM, KNN, and neural network classifiers. Extensive experiments showed that SqueezeNet‐ResNet two‐layer feature fusion with a neural network consistently achieved high accuracy, precision, recall, and F1 score. A three‐layer fusion model with SqueezeNet, ResNet, and MobileNet architectures and a neural network had near‐perfect classification metrics. Results show that multi‐backbone feature evaluation improves classification performance over single‐modality or single‐model baselines. Neural network classifiers outperformed SVM and KNN in all configurations, demonstrating their ability to model complex nonlinear relationships in deep feature representations. For greater clinical impact, this framework may be extended to other tumour types and imaging modalities.