Four Class Brain Tumor Classification in MRI Using Multi Scale Attention Guided Feature Fusion Network
Year: 2026
Type: conference-paper
Abstract: The classification of the type of brain tumor using magnetic resonance imaging is still not accurate because of high inter-class similarity, heterogeneous tumor appearance, and variability presented by acquiring multi-source data. Most of the available methods are based on single-scale feature extraction, and thus have the potential of denying the capacity to identify both fine-scale local features and high-level contextual patterns necessary to make dependable differentiation. To overcome these drawbacks, the present study suggests MSAF-Net, which is a multi-scale attention fusion network that combines parallel convolutional branches with 3x3, 5x5, and 7x7 kernel sizes, and then residual learning, efficient channel attention, and attention-based feature fusion with a support vector machine as the final classifier. The model is trained and tested on a curated multi-source dataset with figshare, SARTAJ, and Br35H datasets, where label errors in the glioma class are fixed and the no tumor class is obtained in Br35H, making the dataset balancing to 5600 training and 1600 test images. A definite train-test split is utilized, so as to guarantee a constant protocol of evaluation devoid of cross-validation. The accuracy and precision of the proposed approach are 99.27 and 99.27, respectively, and the recall and F1-score are 99.16 and 99.21, respectively, with the two phenomena being glioma and meningioma and almost perfect separation between no tumor and the two mentioned phenomena. These findings suggest that multi-scale attention-guided fusion can be used to give a stable and discriminative representation of brain tumor classification in controlled experimental environments.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7214766210