基于BiSeNet V2模型在前路坐骨神经阻滞超声图像分割中的应用研究

Application of BiSeNet V2 Model in Ultrasound Image Segmentation for Anterior Approach Sciatic Nerve Block

摘要:

目的:探讨基于双分支语义分割网络(BiSeNet V2)的人工智能模型在超声引导下前路坐骨神经阻滞图像分割中的应用效果,为神经阻滞定位提供智能化辅助方法。方法:前瞻性收集2025年3月~6月兰州大学第二临床医学院接受超声引导下前路坐骨神经阻滞的患者100例,共获得超声图像3000张。由两名具有丰富经验的麻醉医师采用 ITK-SNAP 软件对神经区域进行人工标注。将采集数据按6∶2∶2比例随机分为训练集、验证集与测试集。模型以双分支结构的 BiSeNet V2 为基础,在 PyTorch 平台上构建并训练,评价指标包括平均交并比(Mean Intersection over Union,mIoU)、平均骰子系数(Mean Dice Coefficient ,mDice)及总体准确率。结果:模型在测试集的mIoU 和 mDice分别为 0.671 和 0.801,总体准确率为 98.7%。经过5折交叉验证的IoU中位数为0.973,模型在不同数据划分下表现稳定。结论:基于 BiSeNet V2 的超声图像分割模型能够在保证实时分割速度的同时实现前路坐骨神经的高精度识别,具备较好的临床应用前景。

Abstract:

Objective:To investigate the application effect of an artificial intelligence model based on the dual-branch semantic segmentation network (BiSeNet V2) in segmenting ultrasound images for anterior approach sciatic nerve blocks, and to provide an intelligent auxiliary method for nerve block localization.Methods:Prospectively, 100 patients undergoing ultrasound-guided anterior sciatic nerve blocks at the Second Clinical Medical College of Lanzhou University were enrolled from March to June 2025, yielding a total of 3000 ultrasound images. Two experienced anesthesiologists manually annotated the nerve regions using ITK-SNAP software. The collected data were randomly divided into training, validation, and test sets in a 6:2:2 ratio. A model based on the dual-branch BiSeNet V2 architecture was built and trained on the PyTorch platform. Evaluation metrics included the Mean Intersection over Union (mIoU), Mean Dice Coefficient (mDice), and overall accuracy.Results:The model achieved an mIoU of 0.671 and an mDice of 0.801 on the test set, with an overall accuracy of 98.7%. The median Intersection over Union (IoU) after 5-fold cross-validation was 0.973 (interquartile range), indicating stable model performance across different data splits.Conclusion:The ultrasound image segmentation model based on BiSeNet V2 can achieve high-precision identification of the sciatic nerve via the anterior approach while ensuring real-time segmentation speed, demonstrating good potential for clinical application.