A frequency selection network for medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Current medical image segmentation methods often struggle with accuracy due to limitations in spatial domain processing, inadequate integration of frequency information, and semantic gaps between features.
• These limitations lead to suboptimal segmentation results, particularly for complex lesions.

Purpose of the Study:

• To introduce a novel Frequency Selection Segmentation Network (FSSN) for more accurate medical lesion segmentation.
• To enhance feature interaction and mitigate semantic gaps by effectively fusing spatial and frequency domain information.

Main Methods:

• The proposed FSSN integrates local spatial features with global frequency information using a Global-Local Feature Aggregation Module (GLAM).
• A Feature Filter Module (FFM) is introduced to manage semantic gaps during cross-level feature fusion, enabling selective preservation of relevant frequency information.
• Deformable convolution (DC) is employed to enhance the focus on critical local features, especially lesion boundaries.

Main Results:

• The FSSN demonstrated superior lesion segmentation accuracy compared to existing methods on two benchmark datasets.
• The network achieved better objective evaluation metrics and subjective visual effects.
• The proposed method requires fewer parameters and exhibits lower computational complexity.

Conclusions:

• The FSSN effectively fuses multi-scale local spatial features and global frequency information for improved medical image segmentation.
• The network successfully addresses semantic gaps and enhances the utilization of pertinent local features, leading to more precise lesion identification.
• FSSN offers a computationally efficient and accurate solution for medical image segmentation tasks.