High frequency edge network for accurate cardiac structure segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate cardiac structure segmentation is crucial for diagnosing cardiovascular diseases.
• Existing methods often struggle with precise boundary delineation due to loss of high-frequency anatomical information.
• Cardiac magnetic resonance imaging (MRI) provides detailed anatomical views but requires robust segmentation techniques.

Purpose of the Study:

• To develop a novel deep learning framework, High Frequency Edge Network (HF-EdgeNet), for improved cardiac MRI segmentation.
• To explicitly model high-frequency components for enhanced boundary delineation in cardiac images.
• To evaluate the performance of HF-EdgeNet against established segmentation baselines.

Main Methods:

• Developed HF-EdgeNet, a high-frequency driven encoder-decoder framework.
• Incorporated High Frequency Edge Transformer (HF-EdgeT) for high-frequency guidance in self-attention.
• Introduced High Frequency Adaptive module (HF-Adapte) to counteract down-sampling degradation.
• Designed Semantic Edge Bridge (SEB) block for high-frequency semantic re-alignment during decoding.

Main Results:

• HF-EdgeNet demonstrated improved average dice scores on ACDC and M&Ms datasets.
• The framework achieved favorable boundary-related performance compared to strong segmentation baselines.
• Explicitly modeling high-frequency information led to superior segmentation accuracy.

Conclusions:

• High Frequency Edge Network (HF-EdgeNet) effectively leverages high-frequency components for cardiac MRI segmentation.
• The proposed architecture enhances anatomical boundary delineation, crucial for clinical applications.
• Explicit high-frequency modeling is a promising direction for boundary-sensitive medical image segmentation.