Artificial Intelligence-Based Segmentation of Local Anesthetic Spread During Ultrasound-Guided Rectus Sheath Block: An Exploratory Observational Study

IntroductionUltrasound-guided rectus sheath block (RSB) requires local anesthetic spread between the posterior surface of the rectus abdominis muscle and the posterior rectus sheath.Injectate distribution is usually assessed visually, and standardized quantitative assessment methods have not yet been established.Previous artificial intelligence (AI) applications in regional anesthesia have primarily targeted anatomical structures; automatic segmentation of dynamic local anesthetic spread remains underexplored.This exploratory study aimed to develop an AI model for automatic segmentation of local anesthetic spread during RSB and evaluate its performance. Materials and methodsThis prospective, single-center observational study included 40 adults who underwent abdominal surgery and received ultrasound-guided RSB between April 2025 and February 2026.Still images were extracted at 10 frames per second from left-sided RSB videos.Cases were divided equally between training and test datasets.Experienced anesthesiologists manually segmented the local anesthetic distribution area to create manually annotated reference masks.A Mask R-CNN (region-based convolutional neural network) model implemented in Detectron2 was trained on the training dataset.Predicted masks were compared with reference masks.The primary outcome was the case-level mean Dice coefficient; secondary outcomes were case-level mean sensitivity and precision.Differences in the Dice coefficient among area groups were assessed using a linear mixed-effects model with case as a random effect, followed by pairwise comparisons of estimated marginal means with Holm adjustment. ResultsThe training and test datasets contained 3,808 and 4,218 still images from 20 cases each, respectively.In the test dataset, the median case-level mean Dice coefficient, sensitivity, and precision were 0.822, 0.867, and 0.874, with interquartile ranges (IQRs) of 0.662-0.865,0.703-0.929,and 0.796-0.911,respectively.The Dice coefficient differed significantly among area groups (p<0.001), with significantly higher values in the medium-and large-area groups than in the small-area group and in the large-area group than in the medium-area group after Holm adjustment. ConclusionThe AI model demonstrated promising segmentation performance for the local anesthetic distribution area on ultrasound images obtained during RSB.These findings support the feasibility of AI-based segmentation of dynamically changing injectate spread and provide a basis for further technical and clinical investigation.