Automated Segmentation of Stellate Ganglion Block Region in Ultrasound Images Using Deep Learning Model

BACKGROUND: The stellate ganglion region is densely vascularized and innervated, making the stellate ganglion block (SGB) technically challenging under ultrasound, particularly for beginners. Deep learning can segment complex ultrasound anatomy, but its application to SGB has not been systematically assessed. We developed and validated a multilevel feature fusion UNet (MLF-UNet) to automatically delineate the SGB region on ultrasound, aiming to support accurate needle placement and improve procedural safety. METHODS: In this retrospective study, 370 patients who underwent ultrasound-guided SGB between March 1, 2023 and January 16, 2025 were included. Three expert anesthesiologists jointly annotated 730 videos (2190 images) to produce ground truth. Data were split 9:1 by patient into development and heldout test sets. MLF-UNet was trained and compared with 5 benchmark models using identical pipelines. Test-set performance was evaluated with Dice similarity coefficient (DSC), Intersection over Union (IoU), 95th percentile Hausdorff distance (95HD), and average symmetric surface distance (ASSD). Three blinded experts rated model outputs (0–2 scale) for topological integrity, boundary precision, and background accuracy. For clinical validation and human–machine comparison, 3 additional experts and 3 nonexperts independently delineated SGB regions on the test set; spatial agreement was visualized with heat maps and assessed by Bland–Altman analysis. Metrics (DSC, IoU, 95HD, and ASSD) were compared among MLF-UNet, experts, and nonexperts. RESULTS: MLF-UNet achieved the best test performance: DSC 0.856 (95% confidence interval CI, 0.846–0.865), IoU 0.754 (95% CI, 0.740–0.768), 95HD 3.98 mm (95% CI, 3.44–4.52 mm), and ASSD 1.08 mm (95% CI, 0.99–1.18 mm). Expert ratings favored MLF-UNet over all benchmark models for topological integrity (all P < .001), boundary precision (all P < .001), background accuracy ( P < .01 or P < .001), and total score (all P < .001). Bland–Altman analysis showed a mean segmentation area difference between MLF-UNet and ground truth of −38.1 mm² (limits of agreement −278 to +202 mm²). MLF-UNet outperformed the nonexpert group on region overlap (DSC, IoU; both P < .001) and boundary precision (95HD, ASSD; both P < .001). Compared with experts, MLF-UNet showed no significant difference in overlap (DSC P = .332; IoU P = .125) but had slightly larger boundary precision (95HD and ASSD: both P < .001). CONCLUSIONS: MLFUNet outperforms 5 benchmark models and nonexpert clinicians for automated ultrasound segmentation of the SGB region, achieving expert‑level region overlap with a modest deficit in boundary precision.