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

Accurate scoliosis diagnosis relies on precise spinal curvature measurement, traditionally using radiographic Cobb's angle. Ultrasound imaging offers a radiation-free alternative via ultrasound curve angle (UCA) estimation, but its clinical utility is limited by incomplete anatomical information due to the restricted field of view (FOV) during scanning. This hinders key tasks like segmentation and landmark detection, restricting ultrasound's broader adoption in scoliosis assessment. To address this challenge, we propose an edge-aware outpainting diffusion framework that restores missing spinal anatomy by integrating mask-guided spatial diffusion. Specifically, the model is trained to predict noise between randomly selected target windows and anchor regions using spatially encoded masks. During inference, a dedicated edge-preservation mechanism guides the generation of anatomical structures. In addition to mitigating hallucinations in diffusion-based generation and ensuring perceptual consistency between generated and retained regions, we incorporate a total variation loss to enforce structural smoothness and coherence across the entire output. This approach effectively constrains reconstruction within masked regions, improving the recovery of anatomical features-particularly in cases of complex S-shaped spinal deformities commonly seen in scoliosis ultrasound imaging, where the field of view is inherently limited. Extensive experiments demonstrate our approach achieves the lowest Fréchet Inception Distance (180.97) and highest Inception Score (1.87 $\pm$ 0.12), while improving thoracic and lumbar UCA estimation accuracy by 47.8% and 24.6%, respectively. Thoracic structure detection also increases by 8.1% compared to Swin-Unet. Low KL divergence and Wasserstein distance confirm strong distributional alignment between generated and real anatomy. Overall, our framework enables anatomically consistent outpainting under limited FOV, enhancing ultrasound's reliability for clinical scoliosis assessment.