Accurate segmentation of the serratus anterior muscle by artificial intelligence during ultrasound-guided serratus anterior plane block: an observational study

Accurate ultrasound identification of the serratus anterior muscle (SAM) and its surrounding key anatomical structures is essential for effective and safe serratus anterior plane block (SAPB). In contrast, errors in identification can lead to block failure or serious complications such as pneumothorax. This study aims to develop a novel artificial intelligence (AI) model, SAPB-Net, to accurately and in real time segment anatomical structures in ultrasound images of SAPB. In this study, anesthesiologists used a high-frequency linear ultrasound probe to scan the SAM along the mid-axillary line and obtain video recordings. Still images were extracted from these videos and manually annotated by experienced anesthesiologists to label the SAM, lungs, and ribs. These annotated images were then used to train and develop SAPB-Net through AI deep learning techniques. To evaluate the performance of SAPB-Net in intelligently segmenting the SAM and its surrounding key structures, a statistical comparison with the baseline UNet model was conducted using the Wilcoxon signed-rank test, with a P -value less than 0.05 considered statistically significant. A total of 2310 labeled ultrasound images were obtained from 167 patients. The Intersection over Union (IoU) and Dice coefficient values for the SAM, lungs, and ribs were 76.75% and 86.85%; 86.25% and 92.62%; and 80.23% and 89.03%, respectively. Furthermore, the accuracy in segmenting these structures reached 92.56%, 91.49% and 90.78%, respectively. Boundary and surface distance metrics also indicated close agreement with manual annotations. Compared with the baseline UNet model, SAPB-Net showed statistically significant improvements across all metrics, with mean Intersection over Union (mIoU) increasing by 1.62% to 84.67%, mean pixel accuracy (mPA) increasing by 1.30% to 93.63%, accuracy increasing by 0.51% to 95.54%, and Dice coefficient increasing by 0.59% to 91.57% (all P < 0.001). SAPB-Net achieved a processing speed of 38.40 frames per second (fps), facilitating real-time anatomical identification during ultrasound-guided SAPB. SAPB-Net demonstrates accurate segmentation of the SAM and surrounding structures, with stable quantitative performance and real‑time processing capability, suggesting its potential to facilitate anatomical identification during ultrasound‑guided SAPB. Registered at Chinese Clinical Trial Registry (ChiCTR2300076361) and registration date: 2023–10-07.