FCFDiff-Net: full-conditional feature diffusion embedded network for 3D brain tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Brain tumor segmentation (BraTS) is vital for diagnosis and treatment planning.
• Diffusion models show promise for image segmentation but struggle with noise and blurring.
• Existing methods often produce false positives and negatives in BraTS.

Purpose of the Study:

• To develop a novel neural network for accurate 3D BraTS.
• To address challenges of noise and blurring in brain tumor segmentation.
• To improve robustness and reduce errors in diffusion model-based segmentation.

Main Methods:

• Proposed a full-conditional feature diffusion embedded network (FCFDiff-Net).
• Introduced the full-conditional feature embedding (FCFE) module for comprehensive feature integration.
• Implemented a multi-head attention residual fusion (MHARF) module to refine segmentation by aligning semantic and noise information.

Main Results:

• Achieved high Dice Similarity Coefficient (DSC) scores on BraTS 2020 and 2021 datasets (e.g., 0.926 for whole tumor on BraTS 2021).
• Demonstrated low Hausdorff distance at 95th percentile (HD95) and false positive rates (FPR) across tumor subregions.
• Showcased excellent specificity (0.999) and reduced errors in noisy or ambiguous regions.

Conclusions:

• FCFDiff-Net offers an efficient and robust solution for 3D BraTS.
• The model outperforms existing methods in accuracy and robustness.
• Future work will explore generalization and lightweight model experiments.