GL-Net: A knowledge-guided Gaussian-gated and layered refinement network for 3D MRI segmentation of brain gliomas.
Authors
Affiliations (4)
Affiliations (4)
- School of Mathematics and Statistics, Changchun University of Technology, Changchun, Jilin, China.
- School of Computer Science and Engineering, Changchun University of Technology, Changchun, Jilin, China.
- The First Hospital of Jilin University, Changchun, Jilin, China.
- Department of Radiation Oncology, Mercy Hospital Cancer Center, Oklahoma, Oklahoma, United States of America.
Abstract
Glioblastoma is a highly malignant brain tumor, and accurate lesion segmentation in MRI is essential for diagnosis, treatment planning, and prognosis assessment. This paper proposes a knowledge-guided 3D hybrid Transformer-CNN framework, GL-Net, which integrates prior knowledge through a Gaussian Gating Module (GGM) and a Layered Refinement Module (LRM), together with a novel Edge-Region Voxel Dynamic Weighted Loss Function. These modules collaboratively enhance feature activation, refine label-specific structures, and improve edge delineation, enabling robust segmentation even under limited-sample conditions. The proposed GL-Net was evaluated on the BraTS2019 and BraTS2021 datasets, achieving average Dice Similarity Coefficients (DSC) of 0.877 and 0.913, and Hausdorff Distances (HD) of 1.83 and 1.55, respectively-demonstrating highly competitive performance and a substantial reduction in boundary errors relative to the reported benchmarks of current data-driven approaches. Furthermore, to assess its clinical applicability, VASARI (Visually Accessible Rembrandt Images) feature extraction was performed using both the GL-Net-generated segmentation masks and the ground truth labels on the BraTS2019 dataset for glioblastoma (GBM) diagnosis. The diagnostic performances were nearly identical (GT AUC: 0.954 / GL-Net AUC: 0.949), and the DeLong test (p = 0.99) indicated no statistically significant difference between the two. These results suggest that GL-Net not only achieves highly competitive segmentation accuracy but also produces radiomic features comparable to expert manual annotations, providing complementary evidence of its potential clinical relevance. The proposed framework shows strong clinical potential for precise and consistent glioma delineation, providing valuable support for surgical planning, radiotherapy targeting, and diagnostic decision-making in clinical workflows.