Rethinking Brain Tumor Segmentation From the Frequency Domain Perspective.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-Oncology

Background:

• Accurate brain tumor segmentation, especially for contrast-enhancing regions in MRI, is vital for clinical decisions but challenging for current methods.
• Existing techniques struggle with complex tumor textures and directional variations inherent in MRI data.
• Insufficient analysis of frequency-domain features limits the performance of current brain tumor segmentation models.

Purpose of the Study:

• To develop a novel deep learning framework, HFF-Net, for improved brain tumor segmentation by leveraging frequency-domain analysis.
• To enhance the characterization of tumor regions by decomposing MRI images into frequency components.
• To improve the sensitivity to tumor boundaries and effectively fuse multi-scale tumor features.

Main Methods:

• Proposed the Harmonized Frequency Fusion Network (HFF-Net) for brain tumor segmentation.
• Introduced a Frequency Domain Decomposition (FDD) module to separate MRI images into low and high-frequency components.
• Developed an Adaptive Laplacian Convolution (ALC) module for emphasizing high-frequency details and a Frequency Domain Cross-Attention (FDCA) module for multi-scale feature fusion.

Main Results:

• HFF-Net achieved an average relative improvement of 4.48% in mean Dice scores across major subregions.
• Demonstrated a significant average relative improvement of 7.33% in segmenting contrast-enhancing tumor regions.
• Validated frequency-domain improvements through visualization, theoretical reasoning, and extensive experiments on four public datasets.

Conclusions:

• HFF-Net significantly enhances brain tumor segmentation accuracy, particularly for contrast-enhancing regions, by utilizing frequency-domain information.
• The proposed FDD, ALC, and FDCA modules effectively capture and fuse critical tumor features, addressing limitations of existing methods.
• HFF-Net offers a promising, computationally efficient, and clinically applicable solution for brain tumor segmentation.