HayabusaNet: Hybrid attention-based multiscale fusion CNN for accurate and efficient brain tumor classification in MRI scans
- 1Division of Electrical Engineering and Computer Science, Graduate School of Natural Science and Technology, Kanazawa University, Kakuma Campus, Kanazawa, 920-1192, Japan
- 2Informatics Study Program, School of Computing, Telkom University, Main Campus (Bandung Campus), Jl. Telekomunikasi No. 1, Bandung, 40257, West Java, Indonesia
- 3School of Entrepreneurial and Innovation Studies, College of Transdisciplinary Sciences for Innovation, Kanazawa University, Kakuma Campus, Kanazawa, 920-1192, Japan
Abstract
Brain tumors are a major clinical threat and are associated with high mortality, making accurate and timely diagnosis from magnetic resonance imaging (MRI) essential for treatment planning. Convolutional neural networks (CNNs) have improved automated MRI analysis. However, many existing models remain computationally expensive and do not fully exploit efficient multiscale feature fusion because they rely on predominantly single-branch, sequential feature extraction. Here, we propose HayabusaNet, a compact CNN that uses depthwise separable convolutions to improve efficiency while preserving discriminative capacity. The network employs parallel multiscale branches with varied kernel sizes and progressively increased channel widths to capture fine-grained and coarse tumor patterns. We introduce a tailored multiscale feature-fusion scheme to integrate branch outputs, and we incorporate a hybrid attention module to enhance salient local cues while injecting global contextual dependencies. HayabusaNet is evaluated on two public benchmarks: the multiclass Kaggle Brain Tumor MRI dataset (7,023 images) and the binary Br35H dataset (3,000 images). Model explanations are examined using Gradient-weighted Class Activation Mapping (Grad-CAM) and Grad-CAM++. On the Kaggle dataset, HayabusaNet achieves 98.86% accuracy, 98.84% precision, 98.78% recall, and a 98.81% F1-score. On Br35H, it attains 98.70% for accuracy, precision, recall, and F1-score. Compared with other deep CNNs, the proposed model improves computational efficiency through fewer parameters, lower FLOPs, and faster inference, while maintaining strong classification performance. In conclusion, these results show that combining multiscale feature fusion, efficient convolutions, and hybrid attention can improve the accuracy–efficiency balance for brain MRI classification. The framework also supports interpretation through saliency visualizations.
Related
- Feature-aware attention convolutional neural network for efficient pulmonary fibrosis classification from chest CT images with FPGA deployment
- TiLCA-Net: nuclei segmentation and classification in pathological images via task-specialized attention and efficient decoding
- EPRA U-Net: an efficient pyramid residual attention framework for accurate infarct segmentation in diffusion-weighted MRI.
- An efficient brain tumor classification system with accurate recognition and complementary downsampling blocks
- Multi-scale convolution with efficient channel attention improves benign–malignant classification of spinal tumors on T2-weighted fat-suppressed MRI
BibTeX
@article{prayogo2026hayabusanet,
title = {{HayabusaNet: Hybrid attention-based multiscale fusion CNN for accurate and efficient brain tumor classification in MRI scans}},
author = {Rizal Dwi Prayogo and Siti Amatullah Karimah and Hidetaka Nambo},
journal = {Expert Systems with Applications},
year = {2026},
month = jun,
doi = {10.1016/j.eswa.2026.133135},
url = {https://doi.org/10.1016/j.eswa.2026.133135},
note = {CC BY},
}
RIS
TY - JOUR
TI - HayabusaNet: Hybrid attention-based multiscale fusion CNN for accurate and efficient brain tumor classification in MRI scans
AU - Rizal Dwi Prayogo
AU - Siti Amatullah Karimah
AU - Hidetaka Nambo
PY - 2026
DA - 2026/06/05
JO - Expert Systems with Applications
DO - 10.1016/j.eswa.2026.133135
UR - https://doi.org/10.1016/j.eswa.2026.133135
AB - Brain tumors are a major clinical threat and are associated with high mortality, making accurate and timely diagnosis from magnetic resonance imaging (MRI) essential for treatment planning. Convolutional neural networks (CNNs) have improved automated MRI analysis. However, many existing models remain computationally expensive and do not fully exploit efficient multiscale feature fusion because they rely on predominantly single-branch, sequential feature extraction. Here, we propose HayabusaNet, a compact CNN that uses depthwise separable convolutions to improve efficiency while preserving discriminative capacity. The network employs parallel multiscale branches with varied kernel sizes and progressively increased channel widths to capture fine-grained and coarse tumor patterns. We introduce a tailored multiscale feature-fusion scheme to integrate branch outputs, and we incorporate a hybrid attention module to enhance salient local cues while injecting global contextual dependencies. HayabusaNet is evaluated on two public benchmarks: the multiclass Kaggle Brain Tumor MRI dataset (7,023 images) and the binary Br35H dataset (3,000 images). Model explanations are examined using Gradient-weighted Class Activation Mapping (Grad-CAM) and Grad-CAM++. On the Kaggle dataset, HayabusaNet achieves 98.86% accuracy, 98.84% precision, 98.78% recall, and a 98.81% F1-score. On Br35H, it attains 98.70% for accuracy, precision, recall, and F1-score. Compared with other deep CNNs, the proposed model improves computational efficiency through fewer parameters, lower FLOPs, and faster inference, while maintaining strong classification performance. In conclusion, these results show that combining multiscale feature fusion, efficient convolutions, and hybrid attention can improve the accuracy–efficiency balance for brain MRI classification. The framework also supports interpretation through saliency visualizations.
SN - 0957-4174
SN - 1873-6793
ER -