DGFI-Net: dual-branch guided feature interaction network for brain tumor segmentation

Introduction Accurate brain tumor segmentation is essential for clinical diagnosis and treatment planning. However, the complex anatomical structures, blurred tumor boundaries, and large inter-patient variability pose significant challenges to existing segmentation models. To address these issues, we propose DGFI-Net, a dual-branch guided feature interaction network. Methods DGFI-Net introduces an auxiliary branch to guide the main branch’s feature learning via hierarchical interaction. Specifically, the main encoding branch is equipped with an efficient context refinement block (ECRB) to capture long-range contextual dependencies. In parallel, the auxiliary encoding branch employs an attention-guided feature refinement block (AGFRB) to emphasize salient tumor regions. At the bottleneck stage of the main encoding branch, we further introduce a lightweight reinforcement module (LRM) to strengthen high-level semantic representations. During the decoding process, the auxiliary branch continuously guides the main branch by transferring both encoding and decoding feature. Results and Discussion Extensive experiments on three public brain tumor datasets (BrainTumor1, BrainTumor2, and BrainTumor3) and a pituitary adenoma dataset collected from Quzhou People’s Hospital demonstrate the superiority of DGFI-Net. The proposed method achieves Dice scores of 0.8402, 0.9011, 0.8040, and 0.9117 on the four datasets, with corresponding Mcc values of 0.8387, 0.8984, 0.8021, and 0.9100, and Jaccard values of 0.7318, 0.8213, 0.6899, and 0.8388. Moreover, comprehensive ablation studies are conducted to verify the effectiveness of each proposed component. These results validate the superiority of the proposed guided dual-branch interaction paradigm for complex medical image segmentation tasks.