Explore the scholarly record.

Abstract

Abstract Background Glioma is a malignant primary brain tumor arising from glial cells, often resulting in fatal outcomes without timely treatment. Accurate and reliable tumor segmentation is essential for clinical decision-making and therapeutic evaluation. This study aims to develop an efficient and precise framework that improves boundary delineation and reduces manual correction requirements. Methods We propose an attention-guided dual-encoder network that integrates a Transformer-based encoder and a CNN-based encoder to jointly capture global contextual information and local spatial details. An attention-guided feature fusion module is introduced to facilitate adaptive integration of complementary cross-branch representations. In the Transformer branch, a sequence-reduction ratio is adopted from existing efficient Transformer designs to reduce the computational cost of self-attention. The model was evaluated on the internal test set ( n = 227), longitudinal scans ( n = 118), and an external dataset ( n = 36). Results AGDE-Net achieved outstanding segmentation performance, with 3D Dice scores of 0.8382 and 0.8044 in the internal testing set, 0.8449 and 0.8374 on longitudinal scans, and 0.7935 and 0.7798 on the external dataset for the tumor and edema regions, respectively. The model also demonstrated excellent computational efficiency, requiring only 10.027 M parameters and 5.943 G FLOPs, with an inference time of 4.91 ms and a memory footprint of 68.05 MB for a single-slice input comprising all four MRI sequences. Conclusion The AGDE-Net offers accurate segmentation of glioma and edema with low computational overhead and fast inference. Its superior accuracy and efficiency highlight its potential for integration into clinical workflows to support diagnosis and treatment planning.