Related Experiment Videos
Knowledge-guided brain tumor segmentation via synchronized visual-semantic-topological prior fusion
1School of Software, Yunnan University, Chenggong Campus, Kunming, Yunnan Province, 650500, China. mingdazhang@ieee.org.
BMC Medical Imaging
|June 4, 2026
Summary
This study introduces a novel framework for brain tumor segmentation, enhancing accuracy by integrating medical knowledge. The Synchronized Tri-modal Prior Fusion (STPF) method improves delineation in challenging boundary regions.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computational Biology
Background:
- Accurate brain tumor segmentation from multi-sequence MRI is crucial for treatment planning.
- Current deep learning methods struggle with ambiguous boundaries due to reliance on visual features and lack of domain knowledge integration.
- Explicitly incorporating medical knowledge like anatomical semantics and geometric topology can improve segmentation performance.
Purpose of the Study:
- To propose a knowledge-guided framework, Synchronized Tri-modal Prior Fusion (STPF), for improved brain tumor segmentation.
- To integrate heterogeneous medical knowledge priors: pathology-driven differential features, unsupervised semantic descriptions, and geometric constraints.
- To enhance segmentation accuracy, especially in ambiguous boundary regions, and demonstrate the framework's stability and generalization capability.
Main Methods:
- Developed the Synchronized Tri-modal Prior Fusion (STPF) framework.
- Integrated three knowledge priors: differential features (T1ce-T1, T2-FLAIR, T1/T2), semantic descriptions, and geometric constraints via persistent homology.
- Employed a dual-level fusion architecture with dynamic prior weighting and nested output heads for hierarchical constraints (ET⊆TC⊆WT).
Main Results:
- STPF achieved a mean Dice coefficient of 0.868 on the BraTS 2020 dataset, outperforming the best baseline by 2.6%.
- Five-fold cross-validation demonstrated stable performance with coefficients of variation between 0.23% and 0.33%.
- Ablation studies revealed performance degradations of 2.8% and 3.5% upon removal of topological and semantic priors, respectively.
Conclusions:
- The STPF framework effectively integrates medical domain knowledge, improving brain tumor segmentation accuracy.
- Explicit integration of anatomical semantics and geometric constraints enhances performance in ambiguous boundary regions.
- STPF shows promising generalization capability and potential for clinical deployment in neuro-oncology.