Deep Learning for QL Block Ultrasound: Good for Landmarks, Struggles with Injection Plane

A nnU-Net model for anterior quadratus lumborum block ultrasound segmentation achieved Dice scores of 0.91 for vertebral body and 0.86 for psoas, but only 0.38 for the injection plane (precision 0.45, recall 0.38).

• Mean Dice across all six classes was 0.62 for nnU-Net vs 0.50 for a manually tuned U-Net.
• Failure analysis: injection plane detection was high but boundary delineation imprecise; posterior renal fascia was often missed but accurate when detected.
• No external validation was performed; further development is needed before clinical application.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.