Related Experiment Video
Updated: Aug 8, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
DBENet: Dual-Branch Encoder Network for brain MRI lesion segmentation
Qiang Zhao1, Zhaohui Zhang2, Tinghua Cao3
1Department of Radiology, People's Hospital of Chongqing Liang Jiang New Area, Chongqing, China.
Frontiers in Neurology
|August 7, 2026
Summary
DBENet, a novel Dual-Branch Encoder Network, enhances brain lesion segmentation in MRI by fusing spatial and frequency features with Segment Anything Model prompts. It achieves superior accuracy and localization, outperforming existing methods.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Accurate brain lesion segmentation in Magnetic Resonance Imaging (MRI) is crucial but challenging due to lesion heterogeneity, scale variations, and unclear boundaries.
- Existing methods struggle to effectively capture diverse lesion characteristics and refine segmentation details.
Purpose of the Study:
- To introduce DBENet, a Dual-Branch Encoder Network designed for improved brain MRI lesion segmentation.
- To leverage complementary spatial and frequency domain features for enhanced lesion representation.
- To integrate Segment Anything Model (SAM) prompt embeddings for context-aware segmentation and boundary refinement.
Main Methods:
- DBENet employs parallel spatial and frequency branches to extract complementary lesion features.
- A Spatial and Frequency Fusion (SFF) module integrates these features into a hybrid representation.
- Lesion-aware prompts encoded by SAM are fused with the hybrid representation using a Multi-scale Attention Fusion (MAF) module for contextual modeling and boundary refinement.
Main Results:
- DBENet achieved state-of-the-art performance on the ISLES 2022 and BraTS 2018 datasets, with Dice scores of 0.8621 and 0.8261, respectively.
- The network demonstrated superior lesion localization and boundary delineation compared to existing approaches.
- Ablation studies validated the effectiveness of the SFF and MAF modules in improving segmentation.
Conclusions:
- DBENet offers a significant advancement in brain MRI lesion segmentation, providing high accuracy and improved localization.
- The model presents a practical solution with a favorable balance between accuracy and computational cost.
- Future research will focus on 3D segmentation, multimodal learning, and foundation model adaptation for broader applicability.
