DGFI-Net: dual-branch guided feature interaction network for brain tumor segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Computational Neuroscience

Background:

• Accurate brain tumor segmentation is critical for effective clinical diagnosis and treatment planning.
• Existing segmentation models face challenges due to complex anatomy, blurred boundaries, and patient variability.
• Advanced deep learning methods are needed to overcome these segmentation difficulties.

Purpose of the Study:

• To propose DGFI-Net, a dual-branch guided feature interaction network for improved brain tumor segmentation.
• To address the limitations of current models in handling complex anatomical structures and inter-patient variability.
• To enhance the accuracy and robustness of medical image segmentation.

Main Methods:

• Developed DGFI-Net, a dual-branch network with hierarchical feature interaction.
• Incorporated an efficient context refinement block (ECRB) for long-range dependencies in the main branch.
• Utilized an attention-guided feature refinement block (AGFRB) in the auxiliary branch to focus on salient tumor regions.
• Introduced a lightweight reinforcement module (LRM) to enhance high-level semantic representations.

Main Results:

• DGFI-Net achieved superior performance on three public brain tumor datasets and a pituitary adenoma dataset.
• Achieved high Dice scores (e.g., 0.9011 on BrainTumor2) and Mcc values (e.g., 0.8984 on BrainTumor2).
• Ablation studies confirmed the effectiveness of individual components and the guided dual-branch interaction paradigm.

Conclusions:

• DGFI-Net demonstrates significant improvements in brain tumor segmentation accuracy and robustness.
• The guided dual-branch interaction approach is effective for complex medical image segmentation.
• The proposed method holds promise for clinical applications in neuro-oncology.