Regional-aware and sequence-informed multi-decoder network for robust brain glioma segmentation in multi-parametric MRI

Abstract

Accurate segmentation of glioblastoma subregions from multi-parametric MRI is essential for diagnosis, surgical planning, and treatment monitoring in neuro-oncology. However, effective delineation of surrounding non-enhancing FLAIR hyperintensity, non-enhancing tumor core, and enhancing tumor remains challenging due to heterogeneous imaging characteristics. Existing deep learning models often fail to fully exploit the clinical specificity of individual MRI sequences. This work introduces a novel deep learning framework featuring (1) a multi-decoder architecture that independently segments key tumor subregions, (2) a sequence-informed guidance strategy that aligns each decoder with MRI sequences best suited to its diagnostic target, and (3) a modified self-attention mechanism for enhanced feature recalibration. These innovations enable precise, region-specific segmentation while preserving anatomical coherence. On the BraTS 2023 dataset, the proposed method achieved an average Dice similarity coefficient (DSC) of 0.9009 and a 95th percentile Hausdorff distance (HD95) of 6.61 mm, surpassing state-of-the-art approaches—particularly for enhancing tumor delineation. Comprehensive ablation studies confirm the contribution of each component. Validation across four external datasets (BraTS 2020, BraTS Africa, MRBrainS18, and BraTS 2024 post-treatment) demonstrates strong generalizability, with DSC gains up to 4.09 % in the most challenging scenarios. By integrating clinical insight with methodological innovation, this framework offers a robust, generalizable solution for glioblastoma segmentation, supporting improved personalized treatment planning and outcome assessment.