A high-frequency feature guided diffusion model for musculoskeletal ultrasound image segmentation

Area of Science:

• Medical imaging and artificial intelligence
• Biomedical image analysis
• Ultrasound technology

Background:

• Musculoskeletal ultrasound (MSKUS) image segmentation is difficult due to severe speckle noise.
• Speckle noise causes boundary ambiguity, hindering accurate anatomical structure delineation.
• Existing segmentation methods are often unsuitable for clinical application due to low accuracy in noisy MSKUS images.

Purpose of the Study:

• To develop a robust method for MSKUS image segmentation that overcomes limitations of existing approaches.
• To improve the accuracy of anatomical structure boundary delineation in the presence of significant speckle noise.
• To provide a more reliable tool for ultrasound-assisted clinical diagnosis.

Main Methods:

• Proposed a high-frequency feature guided diffusion model named HFGSegDiff for robust MSKUS segmentation.
• Employed a dual-branch parallel feature encoding strategy within a conditional diffusion framework.
• Introduced a High-Frequency Cross Attention Module (HFCAM) and a Multi-Scale Feature Enhancement Module (MSFEM) to integrate high-frequency features and multi-scale information for boundary refinement.

Main Results:

• HFGSegDiff outperformed state-of-the-art methods on two public datasets (MUST and DeepACSA) across multiple segmentation metrics.
• Achieved significant improvements in mean Intersection over Union (mIoU) and reduction in Hausdorff Distance 95% (HD95).
• Demonstrated superior noise robustness, with minimal performance degradation under strong speckle noise compared to other methods.

Conclusions:

• The proposed HFGSegDiff model enables accurate extraction of structural boundaries in noisy ultrasound environments.
• Offers a promising solution for robust and precise segmentation in clinical ultrasound applications.
• Facilitates more reliable ultrasound-assisted diagnosis by improving image segmentation accuracy.