腰椎内镜下手术中人工智能动态图像识别模型的建立及其临床有效性探讨

Establishment of an artificial intelligence dynamic image recognition model for lumbar endoscopic surgery and discussion on its clinical effectiveness

中文关键词:  人工智能  内镜  腰椎  椎间孔成形术  语义分割

英文关键词:Artificial intelligence  Endoscope  Lumbar spine  Foraminoplasty  Semantic segmentation

中文摘要:

  【摘要】 目的：开发一种基于语义分割的卷积神经网络人工智能辨识模型，以实现腰椎内镜下手术中组织结构的动态识别，并通过人机对照试验探讨其临床有效性。方法：收集2020年1月~2022年12月60例经椎间孔入路腰椎内镜下手术的视频资料，提取3000张高质量图像。采用Lable studio软件对骨组织、韧带组织和椎间盘组织进行标注并建立标签，分为训练集（2400张）、测试集（300张）和验证集（300张）。利用生成对抗网络技术扩增图像，基于PaddlePaddle深度学习框架训练得到轻量化模型，训练后根据验证集数据选择最优模型，最后利用测试集数据评估模型效果，并对模型能力进行人机验证。应用精确率、准确率、召回率、交并比（intersection over union，IoU）和Dice相似系数（dice similarity coefficient，DSC）结合混淆矩阵评价模型的精度；绘制受试者工作特征曲线（receiver operating characteristic curve，ROC）并计算曲线下面积（area under the curve，AUC）。结果：模型平均准确率达92.72%，平均精确率为80.59%，平均召回率为79.45%，平均IoU为67.28%，平均DSC为79.97%。椎间盘组织、骨组织和韧带组织的AUC分别为0.9779、0.9626和0.9515。人机验证中，高年资医师平均IoU为61.9%，模型为61.0%，中年资医师为59.5%，低年资医师为56.3%，差异无统计学意义（P>0.05）。提示模型精度可以接近高年资医师在分类标签的识别能力。结论：人工智能模型可针对腰椎内镜下重要结构进行标签分类，模型对镜下解剖结构辨识能力与高年资医生的辨识水准相仿。

英文摘要:

  【Abstract】 Objectives: To develop a convolutional neural network artificial intelligence(AI) identification model based on semantic segmentation to realize dynamic recognition of tissue structure during lumbar endoscopic surgery and verify its clinical value through man-machine test. Methods: We collected 60 video clips of lumbar endoscopic surgeries via the transforaminal approach from January 2020 to December 2022, and extracted 3000 high-quality images. Bones, ligaments and intervertebral discs were labeled, after which the images were divided into training set(2400), test set(300), and verification set(300). Using generative adversarial network technology to expand images from segmentation tags, a lightweight model was trained based on PaddlePaddle deep learning framework, and the optimal model was selected according to the verification set data after training. Finally, the model effect was evaluated using the test set data, and the ability of the model was verified in the man-machine test. The model′s performance was evaluated by combining precision, accuracy, recall rate, intersection over union(IoU), and dice similarity coefficient(DSC) with the confusion matrix. The receiver operating characteristic curve(ROC) and the area under the curve(AUC) were plotted. Results: The average accuracy of the model was 92.72%, the average precision was 80.59%, the average recall was 79.45%, the average IoU was 67.28%, and the average DSC was 79.97%. The AUC of intervertebral disc, bone and ligament tissues were 0.9779, 0.9626, and 0.9515, respectively. Taking the average IoU as an example, the man-machine verification results showed that senior physician(61.9%) > model(61.0%) > intermediate physician(59.5%) > junior physician(56.3%), without significant statistical differences between groups in the human-machine verification, suggesting that the model accuracy could approach the recognition ability of senior physicians. Conclusions: The AI model can label and classify the important structure under lumbar endoscopy, and its recognition ability on the anatomical structure under endoscope can reach the level of an senior physician.