130例锥形束CT影像腰椎椎弓根螺钉自动规划的初步分析

目的 利用深度学习技术开发一种基于锥形束CT(CBCT)影像的椎弓根螺钉自动规划系统，验证和评估其在腰椎椎弓根螺钉规划中的有效性和效率。方法 回顾性分析2017年1月至2019年12月行单中心机器人辅助腰椎椎弓根螺钉内固定术中所用CBCT引导影像和椎弓根螺钉规划信息患者130例，其中脊椎分割和椎弓根螺钉自动规划训练100例，算法验证30例。通过Dice系数和定性评估来评价分割结果；基于Ravi分级和Zdichavsky分级评价椎弓根螺钉自动规划结果；基于Babu方法评估椎弓根螺钉上关节突侵犯情况；利用Kappa分析评估观察者间一致性。结果 10例分割定量评估中，Dice值为0.99。30例分割定性评估中，专家对于分割结果的平均可接受率为94.9%,平均满意率为80.5%。共计自动规划了300枚椎弓根螺钉，298枚(99%)为Ravi 1级，300枚(100%)为Zdichavsky IA级。2枚(1%)为Ravi 2级(突破皮质<2 mm),均为向外侧偏出皮质。评估椎弓根螺钉侵犯关节突的情况，300枚螺钉中，294枚(98%)为0级，共计6枚侵犯上关节突，其中3枚(1%)为1级，...

Preliminary exploration of automatic planning of lumbar pedicle screws based on cone-beam CT in 130 cases

Objective To develop an automatic pedicle screw planning system based on cone-beam CT(CBCT)images using deep learning technology, and to evaluate its efficiency and planning efficiency. Methods The CBCT and screw planning data of 130 cases who underwent robot-assisted lumbar pedicle screw fixation from a single academic institution during Jan. 2017 and Dec. 2019 were retrospectively analyzed, including 100 cases for network training for spinal segmentation and automatic planning, and 30 cases for validation. The segmentation results were evaluated using Dice coefficient and qualitative validation. The pedicle screw placement was evaluated using the Ravi and Zdichavsky grading systems. The viocation of superior facet joint was evaluated with Babus method. Interobserver reliability was assessed with Kappa statistics. Results In the quantitative assessment of segmentation in 10 cases, the Dice value was 0.99. In the qualitative evaluation of segmentation in 30 cases, the average acceptance rate of the segmentation results by experts was 94.9%, and the average satisfaction rate was 80.5%. In total, the automatic placed 300 pedicle screws between the L1 and L5 spinal levels, including 300(100%)Zdichavsky 1A, 298(99.0%)Ravi 1, and 2(1.0%)Ravi 2(<2 mm breech), both of which were laterally deviated from the cortex. Of the 300 screws, 294(98%)were Grade 0, 6(2%, all L5)of which had superior facet joint violation, including 3(1%)Grade 1, 2(0.67%)Grade 2, and 1(0.33%)Grade 3. Kappa statistics showed excellent overall agreement between raters(K=0.71-0.94; 90%-96% agreement). In running speed evaluation, the average time required for automatic segmentation was 3.52 seconds, the time required for feature point extraction was 3.81 seconds, and the total average running time was 7.33 seconds. Conclusion The automatic planning system of pedicle screws based on CBCT images shows excellent automatic segmentation and screw planning, which has the potential to aid surgeons in screw planning and improve the navigation workflow of spinal surgical robots.