CAGs-Net: A Novel Adjacent-Context Network With Channel Attention Gate for 3D Brain Tumor Image Segmentation

Accurate brain tumor segmentation is essential for clinical decision-making, yet remains difficult to automate. Key obstacles include the small volume of lesions, their morphological diversity, poorly defined MRI boundaries, and nonuniform intensity profiles. Furthermore, while traditional segmentation approaches often focus on intralayer relevance, they frequently underutilize the rich semantic correlations between features extracted from adjacent network layers. Concurrently, classical attention mechanisms, while effective for highlighting salient regions, often lack explicit mechanisms for directing feature refinement along specific dimensions. To solve these problems, this paper presents CAGs-Net, a novel network that progressively constructs semantic dependencies between neighboring layers in the UNet hierarchy, enabling effective integration of local and global contextual information. Meanwhile, the channel attention gate was embedded within this adjacent-context network. These gates strategically fuse shallow appearance features and deep semantic information, leveraging channel-wise relationships to refine features by recalibrating voxel spatial responses. In addition, the hybrid loss combining generalized dice loss and binary cross-entropy loss was employed to avoid severe class imbalance inherent in lesion segmentation. Therefore, CAGs-Net uniquely combines adjacent-context modeling with channel attention gates to enhance feature refinement, outperforming traditional UNet-based methods, and the experimental results demonstrated that CAGs-Net shows better segmentation performance in comparison with some state-of-the-art methods for brain tumor image segmentation.