RGA Attention: Residual Gaussian Attention for Enhancing CNN Performance in Brain Tumor Classification
Year: 2026
Type: conference-paper
Abstract: Convolutional Neural Networks (CNNs) have become the backbone of modern medical image analysis. In brain tumor classification, precise identification of tumor types is critical for effective diagnosis and treatment planning. Despite their success, conventional CNNs treat all feature representations equally, potentially overlooking discriminative tumor regions. In this study, we propose a novel attention mechanism-RGA Attention (Residual Gaussian Attention) that effectively captures both spatial and channel dependencies while introducing a Gaussian inhibition strategy to refine attention maps. This enables the network to focus on tumor-relevant features while avoiding over-focusing on highly activated areas. To validate its effectiveness, we integrate RGA Attention into four widely used CNN architectures (ResNet50, InceptionV3, DenseNet121, and Xception) and evaluate them on a benchmark brain tumor MRI dataset. Experimental results show that RGA Attention consistently outperforms existing modules, achieving the highest accuracy across all tested architectures.
Institution: FPT University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203727770