Edge Extension for Missing Anatomical Features: A Mask-Guided Spatial Diffusion Framework for Ultrasound Scoliosis...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Scoliosis diagnosis traditionally uses radiographic Cobb's angle, exposing patients to radiation.
• Ultrasound imaging offers a radiation-free alternative for scoliosis assessment via ultrasound curve angle (UCA) estimation.
• Limited field of view (FOV) in ultrasound hinders anatomical detail, impacting segmentation and landmark detection for scoliosis.

Purpose of the Study:

• To develop an AI framework for enhancing ultrasound imaging in scoliosis assessment.
• To restore missing spinal anatomy in limited FOV ultrasound scans using an edge-aware outpainting diffusion model.
• To improve the accuracy and reliability of ultrasound-based scoliosis diagnosis.

Main Methods:

• Proposed an edge-aware outpainting diffusion framework integrating mask-guided spatial diffusion to restore spinal anatomy.
• Trained the model to predict noise between target windows and anchor regions using spatially encoded masks.
• Incorporated an edge-preservation mechanism and total variation loss for anatomically consistent and smooth reconstructions.

Main Results:

• Achieved state-of-the-art performance with the lowest Fréchet Inception Distance (180.97) and highest Inception Score (1.87 ± 0.12).
• Significantly improved thoracic (47.8%) and lumbar (24.6%) UCA estimation accuracy.
• Enhanced thoracic structure detection by 8.1% compared to Swin-Unet, with strong distributional alignment confirmed by low KL divergence and Wasserstein distance.

Conclusions:

• The developed framework enables anatomically consistent outpainting under limited FOV ultrasound imaging.
• This advancement enhances the reliability and clinical utility of ultrasound for scoliosis assessment.
• The AI approach overcomes FOV limitations, improving diagnostic capabilities for complex spinal deformities.