基于深度学习的前路坐骨神经超声图像分割OA

Objective To construct a set of ultrasound image datasets and a deep learning-based ultrasound im-age recognition system in order to explore the recognition method for the sciatic nerve block area in the anteri-or approach region.Methods Ultrasound images of their anterior approach sciatic nerve were collected and manually labeled using ITK-SNAP software,and a dataset was established.The PyTorch deep learning frame-work was used for training data processing and segmentation output of regions of interest.The model perfor-mance was evaluated using the mean intersection over union(mIoU)and mean dice similarity coefficient(mDice)as evaluation metrics.Results A total of 3 000 labeled ultrasound images were used as the dataset,including 1 800 images for the training set,600 for the validation set,and 600 for the test set.For the sciatic nerve in the test set,the mIoU was 0.675,the mDice coefficient was 0.775,the model achieved a mean F1-score of 0.718 and a median of 0.720.A 5-fold cross-validation determined the median of mIoU for the sciatic nerve to be 0.805.Conclusion Based on the deep learning model,favorable results were achieved in the auto-matic identification of the anatomical structure of the sciatic nerve in the anterior approach region.This meth-od enables a safe and precise injection of local anesthetics into the paraneural space,presenting promising clin-ical application value.