Feasibility of Deep Learning Segmentation in Anterior Quadratus Lumborum Ultrasound Imaging: Challenges in Identifying the Injection Plane.

Accurate identification of sonoanatomy is essential for successful anterior quadratus lumborum (QL) block performance. Prior work has addressed anatomically distinct muscular and osseous structures, whereas automated segmentation of the anterior QL block injection plane has not been described. In this single-center prospective imaging study, 82 simulated anterior QL block ultrasound recordings from 42 healthy volunteers yielded 460 annotated frames across six classes, including the anterior QL block injection plane. A 2D nnU-Net model was trained using fivefold cross-validation and tested on 23 unseen clinical images acquired during routine clinical practice. Performance was assessed using Dice similarity coefficients, intersection-over-union (IoU), precision, and recall, and compared with a manually tuned 2D U-Net. The mean Dice score across all classes was 0.62, compared with 0.50 for the U-Net baseline. Performance was highest for the vertebral body (0.91) and psoas major muscle (0.86), moderate for the quadratus lumborum muscle (0.69) and transverse abdominal muscle (0.52), and lower for the posterior renal fascia (0.35) and injection plane (0.38). Precision and recall distinguished two failure modes. The posterior renal fascia was frequently missed but accurate when detected (0.74 and 0.31), whereas the injection plane was localized in nearly all cases but imprecisely delineated (0.45 and 0.38). Deep learning was feasible for segmentation of major anatomical landmarks in anterior QL ultrasound images, but remained limited for the interfascial injection plane, where the difficulty lay in boundary precision rather than detection. Further development and external validation are required before clinical application.