Automated Segmentation of Stellate Ganglion Block Region in Ultrasound Images Using Deep Learning Model
Authors' conclusion
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.
Study context
- Population
- Patients who underwent ultrasound-guided stellate ganglion block (SGB), March 1, 2023 to January 16, 2025730 videos (2190 images)
- Participants
- 370 patients
- Comparison
- MLF-UNet vs 5 benchmark models, experts and nonexperts for ultrasound segmentation of the SGB region
- Benchmark modelsn = 5reference
- Expertsn = 3reference
- Nonexpertsn = 3reference
- Outcome
- Test-set segmentation performance: DSC, IoU, 95th percentile Hausdorff distance (95HD), average symmetric surface distance (ASSD)
Methodological detailsdeep learning model development and validation, patient-level 9:1 split into development and held-out test sets, blinded expert rating, human-machine comparison, Ground truth jointly annotated by 3 expert anesthesiologists
- Deep learning model development and validation
- Patient-level 9:1 split into development and held-out test sets
- Blinded expert rating
- Human-machine comparison
- Ground truth jointly annotated by 3 expert anesthesiologists
Performance
As reported in the abstract
Test-set segmentation performance of MLF-UNet
| Test or model | Dice similarity coefficient (DSC) | Intersection over Union (IoU) | 95HD | Average symmetric surface distance (ASSD) |
|---|---|---|---|---|
| MLF-UNet | 0.856 (95% CI, 0.846-0.865) | 0.754 (95% CI, 0.740-0.768) | 3.98 mm (95% CI, 3.44-4.52 mm) | 1.08 mm (95% CI, 0.99-1.18 mm) |
MLF-UNet
Dice similarity coefficient (DSC)0.856 (95% CI, 0.846-0.865)
Intersection over Union (IoU)0.754 (95% CI, 0.740-0.768)
95HD3.98 mm (95% CI, 3.44-4.52 mm)
Average symmetric surface distance (ASSD)1.08 mm (95% CI, 0.99-1.18 mm)
Further findings
| Outcome | Mean segmentation area difference (mm²) | P value |
|---|---|---|
| Expert ratings (0-2 scale) vs benchmark modelsMLF-UNet vs Benchmark models | all P < .001 | |
| Expert ratings (0-2 scale) vs benchmark modelsMLF-UNet vs Benchmark models · background accuracy favored MLF-UNet | P < 0.01 or P < .001 | |
| Segmentation area agreement with ground truth (Bland-Altman)MLF-UNet vs Ground truth · limits of agreement −278 to +202 mm² | −38.1 | |
| MLF-UNet vs nonexpertsMLF-UNet vs Nonexperts · MLF-UNet better on region overlap (DSC, IoU) | both P < .001 | |
| MLF-UNet vs nonexpertsMLF-UNet vs Nonexperts · MLF-UNet better on boundary precision (95HD, ASSD) | both P < .001 | |
| MLF-UNet vs expertsMLF-UNet vs Experts · no significant difference in overlap | P = 0.332 | |
| MLF-UNet vs expertsMLF-UNet vs Experts · no significant difference in overlap | P = 0.125 | |
| MLF-UNet vs expertsMLF-UNet vs Experts · slightly larger boundary precision values for MLF-UNet | both P < .001 |
- Expert ratings (0-2 scale) vs benchmark models: topological integrity, boundary precision and total score favored MLF-UNet
Original publication
Automated Segmentation of Stellate Ganglion Block Region in Ultrasound Images Using Deep Learning Model. Anesth Analg. Published 16 February 2026. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →