DSAF-Net: A Dynamic Scale-Aware Attention Fusion Network for Multiclass Brain Tumour Classification from MRI Images

• Oluwadare A. ADEBISI

  Author

• Ibiyemi I. ADEAGA

  Author

• Samson O. AYANLADE

  Author

• Khadijah O. LAWAL

  Author

Accurate classification of brain tumours from magnetic resonance imaging (MRI) plays a crucial role in early diagnosis, treatment planning, and clinical decision-making. Although deep learning approaches have advanced automated tumour analysis, many existing models remain constrained by inadequate feature representation, loss of discriminative information, inconsistent generalisation, and growing computational demands. Addressing these challenges requires learning frameworks capable of capturing heterogeneous tumour characteristics while maintaining practical computational efficiency.  This study presents a Dynamic Scale-Aware Attention Fusion Network (DSAF-Net) for multiclass brain tumour classification from MRI images. The proposed architecture integrates residual multi-scale feature extraction, Dynamic Scale-Aware Attention Fusion modules, hybrid spatial-channel attention, and cross-scale feature fusion within a unified framework. Through adaptive weighting and aggregation of complementary feature representations across multiple receptive fields and network depths, the model enhances discriminative feature learning while preserving computational efficiency. Evaluation was conducted using 4,400 MRI images obtained from Kaggle. The proposed framework achieved an accuracy of 98.94%, precision of 98.87%, recall of 98.76%, F1-score of 98.81%, and an area under the receiver operating characteristic curve (AUC) of 99.72%. Comparative analysis demonstrated consistent performance improvements over SqueezeNet, DenseNet121, and EfficientNet-B7. By combining high predictive accuracy with computational practicality, the proposed framework offers a promising direction for the development of adaptive multi-scale attention-based solutions in medical image analysis.

• 31_Adebisi et al.(2026)

Copyright (c) 2026 Oluwadare A. ADEBISI, Ibiyemi I. ADEAGA, Samson O. AYANLADE, Khadijah O. LAWAL (Author)

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.