AdaRFNet: Adaptive receptive field network with Convolution Block Attention Module for brain tumor classification

Accurate brain tumor classification from magnetic resonance imaging (MRI) requires deep learning models capable of capturing multi-scale contextual information while remaining computationally efficient under the constraints of medical imaging. In this work, we propose AdaRFNet, a lightweight convolutional neural network that integrates SKNet-based adaptive receptive field learning with the Convolutional Block Attention Module (CBAM). The proposed architecture employs dynamic kernel selection with multiple receptive field sizes and dual attention mechanisms to adaptively emphasize tumor-relevant features, while maintaining compatibility with limited training data and modest computational resources. Comprehensive ablation studies across multiple architectural variants and four network depths (AdaRFNet18, AdaRFNet34, AdaRFNet50, and AdaRFNet101) are conducted to analyze the contributions of adaptive receptive field learning and attention mechanisms. Statistical evaluation using paired t-tests across 5-fold cross-validation demonstrates that the AdaRFNet models achieve consistent and statistically significant performance improvements over baseline models. The model's generalization capability is validated on two independent benchmark datasets. AdaRFNet18 achieves a test accuracy of 99.24% on the Kaggle dataset, while AdaRFNet34 achieves a test accuracy of 98.14% on the Figshare dataset, consistently improving precision, recall, and F1-score across both benchmarks. Grad-CAM visualizations further confirm that the model focuses on clinically relevant tumor regions. These results demonstrate that AdaRFNet achieves strong classification performance, robustness, and generalization across independent benchmarks, establishing it as an efficient and reliable solution for brain tumor classification.