摘要
目的 建立基于深度学习的人工智能术式选择系统,探讨其指导腰椎内镜手术选择的效果.方法 收集天津市天津医院2018年1月至2023年6月1 110例接受经皮经椎间孔内镜、804例接受经皮椎板间入路内镜、923例接受可动式椎间盘镜及623例接受单侧双通道内镜手术患者的一般资料,包括术前及术后12个月时的腿痛、腰痛视觉模拟评分(visual analogue scale,VAS)、Oswestry 功能障碍指数(Oswestry disability index,ODI)及 MacNab标准.使用随机数字表法按照4∶1将患者分为训练数据集(2 768例)和测试数据集(692例).将患者的临床症状、体征、影像学多模态数据输入深度学习模型,对影像学数据感兴趣区域进行标注,模型分为椎间盘定位模块、是否需要手术的判别模块和术式推荐模块,采用以U-Net分割模型、ResNet为骨干的卷积神经网络与最终采用的术式进行整合,建立训练数据库.计算测试集各模块的准确率和召回率.结果 所有患者与术后随访12个月时相比,接受经皮经椎间孔内镜、经皮椎板间入路内镜、可动式椎间盘镜和单侧双通道内镜手术患者的腿痛VAS评分,分别从术前的(7.69±0.80)、(7.82±0.88)、(7.62±0.69)和(7.56±1.00)分降至(1.44±1.09)、(1.35±0.82)、(1.51±1.08)和(1.43±0.91)分;腰痛 VAS 评分分别从术前的(5.73±0.83)、(6.17±0.99)、(6.11±0.88)和(6.46±0.95)分降至(0.93±0.75)、(1.01±0.67)、(1.40±0.72)和(1.27±0.70)分;ODI 分别从术前的39.91％±4.50％、40.05％±8.05％、47.08％±9.50％和 44.43％±4.71％降至 5.77％±2.22％、6.05％±2.31％、8.51％±2.16％和 9.51％±3.70％,差异均有统计学意义(P＜0.05).MacNab标准评价:优1 637例、良1 585例、可235例、差3例,优良率为93.12％(3222/3 460).在深度学习模型中,训练集输入2768例患者的多模态数据进行深度学习,形成手术识别和术式推荐系统,测试集输入692例患者术前资料,与最终采用的术式相对比.在椎间盘定位模块中,患者腰椎间盘的位置与名称的准确率为97.1％(672/692).在是否需要手术的判别模块中,以椎间盘为单位的判断准确率为94.80％(3 280/3 460),召回率为91.9％(636/692);以患者为单位的判断准确率为91.9％(636/692).在术式推荐的模块中,以椎间盘为单位的系统推荐术式准确率为89.5％(569/636),以患者为单位的系统推荐术式准确率为82.2％(569/692).结论 成功建立的基于深度学习的人工智能术式选择系统可有效地指导腰椎内镜手术的选择.
Abstract
Objective To establish an artificial intelligence-based surgical selection system utilizing deep learning to as-sist in the decision-making process for lumbar endoscopic surgery.Methods General data of 1,110 patients who underwent per-cutaneous transforaminal endoscopic discectomy,804 patients who underwent percutaneous interlaminar endoscopic discectomy,923 patients who underwent mobile microendoscopic discectomy and 623 patients who underwent unilateral biportal endoscopic in Tianjin Hospital from January 2018 to June 2023 were included in the study.Clinical outcomes were assessed using the visual ana-logue scale(VAS)for leg and back pain,the Oswestry disability index(ODI),and MacNab criteria both before surgery and 12 months postoperatively.Using a random number table method,patients were divided into a training dataset(2,768 cases)and a test dataset(692 cases)at a ratio of 4∶1.Patient clinical symptoms,physical signs,and multi-modal imaging data were input into a deep learning model.This model was structured into three main modules:intervertebral disc detection,surgical necessity identifi-cation,and surgical recommendation.The final surgical method was determined using a convolutional neural network incorporat-ing U-Net for segmentation and ResNet for classification.The accuracy and recall rates of each module were evaluated using the test dataset.Results Compared to preoperative values,all patients showed significant improvements at the 12-month postopera-tive follow-up.For patients who underwent percutaneous transforaminal endoscopic discectomy,percutaneous interlaminar endo-scopic discectomy,mobile microendoscopic discectomy,and unilateral biportal endoscopic surgery,the VAS scores for leg pain de-creased from 7.69±0.80,7.82±0.88,7.62±0.69,and 7.56±1.00 preoperatively to 1.44±1.09,1.35±0.82,1.51±1.08,and 1.43±0.91 postoperatively.Similarly,the VAS scores for back pain decreased from 5.73±0.83,6.17±0.99,6.11±0.88,and 6.46±0.95 to 0.93±0.75,1.01±0.67,1.40±0.72,and 1.27±0.70,respectively.Additionally,the ODI significantly decreased from 39.91％±4.50％,40.05％±8.05％,47.08％±9.50％,and 44.43％±4.71％preoperatively to 5.77％±2.22％,6.05％±2.31％,8.51％±2.16％,and 9.51％±3.70％postoperatively,with all differences being statistically significant(P＜0.05).The excellent rate according to the MacNab criteria was 93.12％(3,222/3,460).In the deep learning model,the multi-modal data of 2,768 patients were input in the training set for deep learning to form a surgical identification and operation recommendation system,and the preoperative data of 692 patients were input in the test set to compare with the final operation method.In the intervertebral disc location module,the accuracy of location and designation of the five lumbar intervertebral discs was 97.1％(672/692).In the module of intervertebral disc need for surgery,the accuracy was 94.8％(3,280/3,460)and the recall rate was 91.9％(636/692).As for patients,the accuracy rate was 91.9％(636/692).In the operation recommendation module,the accuracy rate of operation recommendation based on inter-vertebral disc was 89.5％(569/636),and the accuracy rate of surgical recommendation based on patient was 82.2％(569/692).Conclusion In this study,an artificial intelligent surgical procedures selection system based on deep learning was established,which could effectively integrate relevant data and accurately guide the selection of lumbar endoscopic surgery.
基金项目
天津市卫生健康科技项目(TJWJ2023XK023)
天津市卫生健康科技项目(TJWJ2023QN054)
天津市自然科学基金(23JCYBJC01400)
维普
万方数据