Kfd-net: a knowledge fusion decision method for post-processing brain glioma MRI segmentation

The automatic segmentation of brain glioma in MRI images is of great significance for clinical diagnosis and treatment planning. However, achieving precise segmentation requires effective post-processing of the segmentation results. Current post-processing methods fail to differentiate processing based on the glioma category, limiting the improvement of MRI segmentation accuracy. This paper proposes a novel knowledge fusion decision method for post-processing brain glioma MRI segmentation. The method takes grading information and the area ratio from the initial segmentation as input, performs fuzzy reasoning based on formulated rules, and generates decision coefficients for different segmentation regions. To address class imbalance in the segmentation network, a Boundary Region Voxel Dynamic Weighted Loss Function is introduced. On the BraTS2019 validation set, our method achieves DSC values of 0.756, 0.990, and 0.805 for ET, WT, and TC regions, respectively, along with HD values of 4.02mm, 10.73mm, and 9.52mm. Compared to state-of-the-art methods, our proposed approach demonstrates superior segmentation performance. Validation on the BraTS2020 dataset further confirms the stability and reliability of our method.