ConvBiFuseNet: a parallel fusion model with routing attention for MRI brain tumor classification

Although convolutional neural networks (CNNs) have made significant progress in medical image processing, the standard CNN model still has some drawbacks. Firstly, local feature extraction in CNN networks lacks targeted adaptive attention, potentially resulting in ineffective feature information being emphasized. Secondly, the model fails to effectively capture dependencies among long-range features, leading to suboptimal discriminative feature representations for each semantic category. To address these limitations, this paper proposes the ConvBraNet model with adaptive attention capability, introducing a Bi-level Routing Attention (BRA) and Global Response Normalization (GRN). BRA enhances the interdependence of channel mappings by employing an adaptive attention mechanism with routing, allowing irrelevant feature information to be ignored. Ultimately, we integrated ConvBraNet with Biformer, naming the resulting model ConvBiFuseNet. This fusion model combines ConvBraNet’s excellent local feature extraction performance with Biformer’s ability to capture global feature context information, utilizing a parallel integrated staged interaction learning approach. This approach effectively integrates local information with corresponding global dependencies. The proposed model’s performance is evaluated on three different brain tumor datasets for classifying brain tumor MRI 2D slice images. Notably, on the 2D T1-weighted CE-MRI dataset, the model achieves an accuracy of 98.36%, a recall of 98.53%, a specificity of 99.24%, and an F1-Score of 98.15%. Furthermore, the model’s generalization ability is verified on other datasets, yielding satisfactory results. Comparison and ablation experiments confirm the importance of the attention mechanism model with routing and the effectiveness of the proposed fusion model architecture.