AI Research Paper Proposes Dual Backbone Network for Brain Tumor Classification with 99.24% Accuracy
A new deep learning model called DB-FGA-Net demonstrates exceptional performance in multi-class brain tumor classification, achieving 99.24% accuracy on the 7K-DS dataset without requiring data augmentation. The system integrates VGG16 and Xception backbones with a Frequency-Gated Attention Block to capture both local and global features, showing robustness across variably sized datasets. For clinical transparency, Grad-CAM visualization highlights the specific tumor regions influencing predictions, addressing interpretability concerns in medical applications. This approach advances neuro-oncology diagnostics by combining high accuracy with explainable AI techniques.
Key facts
- DB-FGA-Net achieves 99.24% accuracy on 7K-DS dataset for 4-class brain tumor classification
- Model integrates VGG16 and Xception backbones with Frequency-Gated Attention Block
- System operates without data augmentation, demonstrating strong generalization
- Grad-CAM visualization provides clinical interpretability of predictions
- Additional performance metrics include 98.68% and 99.85% accuracy in other settings
- Research addresses trust limitations in clinical deep learning applications
- Paper published on arXiv with identifier 2510.20299v3
- Method captures complementary local and global tumor features
Entities
Institutions
- arXiv