当前位置：首页 > news >正文

基于形状补全和形态测量描述符的腓骨游离皮瓣下颌骨重建自动规划|文献速递-深度学习医疗AI最新文献

news 2025/7/6 18:41:57

Title

题目

Automated planning of mandible reconstruction with fibula free flap basedon shape completion and morphometric descriptors

基于形状补全和形态测量描述符的腓骨游离皮瓣下颌骨重建自动规划

文献速递介绍

因创伤、骨髓炎和肿瘤而接受下颌骨节段切除术的患者，常常会出现下颌骨连续性缺失的情况，这导致他们在面部外观、咀嚼和发音方面遇到困难（洛佩斯 - 阿尔卡斯等人，2010）。带血管的自体骨移植，其中腓骨是一种常用的供骨类型，为修复和重建下颌骨缺损提供了有效的解决方案（洛夫斯特兰德等人，2018）。然而，腓骨的自然形状与患者下颌骨的形状差异显著。因此，外科医生会对腓骨进行截骨操作，将其分割成几个节段，然后重新摆放这些腓骨节段，以达到期望的与下颌骨轮廓的相似度（马等人，2022；赵等人，2012）。在这一外科手术过程中，有效地保持面部美观是口腔颌面外科医生面临的最大挑战之一（库马尔等人，2016；张等人，2020）。传统上，在手术前通常会进行手术规划，使用计算机辅助手术（CAS）技术来模拟腓骨节段的预期位置（赖等人，2022；托托等人，2015；诺比斯等人，2020），这通常包括：（i）使用镜像法模拟完整下颌骨的形态；（ii）进行虚拟截骨，将腓骨切割成几个节段；（iii）拼接腓骨节段以匹配下颌骨的形状。然而，受肿瘤或创伤影响的患者下颌骨在手术前可能会发生形状变化。在计算机辅助手术中，通常采用镜像法来模拟完整的形态。这需要通过视觉确定下颌骨的中轴，将完整的一侧与受影响的一侧分开。为了模拟患病前下颌骨的完整形状，将完整的一侧沿着矢状面进行镜像（哈尼等人，2019；斯特林·克雷格等人，2015）。然后，外科医生反复调整腓骨截骨平面的位置和角度，并将腓骨节段拼接成期望的下颌骨形状。尽管计算机辅助手术为外科医生提供了视觉上的帮助，但明显缺乏确保达到期望术后面部外观的客观标准。由于下颌骨存在内在的发育和生长差异，通过镜像法恢复下颌骨完整形态的准确性和可行性无法得到保证，这表明下颌骨并非完全对称（卡普尔等人，2018；纳希德，2018）。此外，这种方法的适用性有限，因为它只适用于缺损未超过下颌骨中线的患者，对于双侧下颌骨缺损的患者则无法进行镜像操作（菲辛格等人，2019）。而且，也缺乏衡量腓骨皮瓣与下颌骨之间形态差异的客观标准（安东尼等人，2011）。目前，外科医生主观地确定腓骨截骨平面的位置和角度，或者通过鼠标输入数值参数来确定。这种规划过程完全依赖于手术团队的主观评估，会因他们的经验和偏好而产生差异（斯莫尔卡等人，2008）。许多研究都集中在使用三维CT图像的下颌骨腓骨皮瓣重建术前规划系统的有效性和精确性上（德·梅斯沙尔克等人，2017）。然而，还没有团队尝试开发完全客观的标准来量化规划过程。这些情况促使我们引入一种全新的、完全自动化的规划方法，该方法可以同时解决决定规划准确性的关键因素：完成缺损下颌骨的形态补全，以及衡量下颌骨与腓骨之间形态差异的标准。在这项研究中，我们提出了一个新颖的两步式框架，用于腓骨游离皮瓣下颌骨重建的术前自动规划。首先，我们采用带有形状先验信息的卷积神经网络来学习下颌骨的整体形状结构。我们将肿瘤切除后的下颌骨缺损部分作为网络输入，以预测缺损部位完整的下颌骨形态。由于下颌骨的形态具有高度的个体性，为了有效地规划手术，恢复因肿瘤或创伤导致的术前下颌骨形态变化至关重要。接下来，我们从二维角度评估预测的缺损下颌骨（(\mathcal{M}{pre})）与当前规划方案之间的形状差异，生成一个目标函数。我们在(\mathcal{M}{pre})中定义这个目标函数的变量和可行域，并研究函数变量与规划解决方案之间的三维空间变换。随后，我们使用一种优化算法来最小化目标函数（蔡等人，2020）。为了证明我们自动规划过程的准确性和可行性，我们进行了一项回顾性研究和一个临床案例分析。我们的主要贡献总结如下：（1）我们提出了一个用于腓骨游离皮瓣（FFF）下颌骨重建术前自动规划的两阶段框架。这个框架将患病前形态预测与基于优化的移植规划相结合，为手术规划提供了一种系统且标准化的方法，同时保留了对个别患者需求的适应性。（2）在第一阶段，我们应用了一种基于深度学习的方法，该方法结合了形状先验信息来预测患病前的下颌骨形态。这种方法显示出能够提供符合解剖学合理性的参考，即使是对于涉及严重缺损或跨越中线畸形的病例也是如此，而这些病例在手动规划中特别具有挑战性。（3）在第二阶段，我们开发了一种基于优化的规划策略，该策略定义并优先考虑影响移植形态和放置的关键参数。这种方法为规划评估建立了客观基础，减少了对人工调整的依赖，并支持手术规划的可重复性。（4）所提出的框架通过一系列实验进行了严格验证。结果表明，该框架优于以前的方法，取得的结果可与经验丰富的外科医生的专业水平相媲美。这证实了所提出方法在下颌骨重建中的准确性和效率。一项涉及下颌骨中段缺损患者的临床实验表明，该框架在临床应用中具有可行性，突出了该方法的临床潜力。

Abatract

摘要

Vascularized fibula free flap (FFF) grafts are frequently used to reconstruct mandibular defects. However,the current planning methods for osteotomy, splicing, and fibula placement present challenges in achievingsatisfactory facial aesthetics and restoring the original morphology of the mandible. In this study, we proposea novel two-step framework for automated preoperative planning in FFF mandibular reconstruction. Theframework is based on mandibular shape completion and morphometric descriptors. Firstly, we utilize a 3Dgenerative model to estimate the entire mandibular geometry by incorporating shape priors and accountingfor partial defect mandibles. Accurately predicting the premorbid morphology of the mandible is crucial fordetermining the surgical plan. Secondly, we introduce new two-dimensional morphometric descriptors to assessthe quantitative difference between the planning scheme and the full morphology of the mandible. We havedesigned intuitive and valid variables specifically designed to describe the planning scheme and constructedan objective function to measure the difference. By optimizing this function, we can achieve the best shapematched 3D planning solution. Through a retrospective study involving 65 real tumor patients, our methodhas exhibited favorable results in both qualitative and quantitative analyses when compared to the plannedresults of experienced clinicians using existing methods. This demonstrates that our method can implement anautomated preoperative planning technique, eliminating subjectivity and achieving user-independent results.Furthermore, we have presented the potential of our automated planning process in a clinical case, highlightingits applicability in clinical settings

带血管的腓骨游离皮瓣（FFF）移植常用于下颌骨缺损的重建。然而，目前在截骨、拼接以及腓骨放置方面的规划方法，在实现令人满意的面部美学效果以及恢复下颌骨原本形态上存在挑战。在这项研究中，我们提出了一个新颖的两步式框架，用于腓骨游离皮瓣下颌骨重建的术前自动规划。该框架基于下颌骨的形状补全和形态测量描述符。首先，我们利用一个三维生成模型，通过纳入形状先验信息并考虑部分缺损的下颌骨来估计整个下颌骨的几何形状。准确预测下颌骨患病前的形态对于确定手术方案至关重要。其次，我们引入了新的二维形态测量描述符，以评估规划方案与下颌骨完整形态之间的量化差异。我们专门设计了直观且有效的变量来描述规划方案，并构建了一个目标函数来衡量这种差异。通过优化这个函数，我们能够实现最佳的形状匹配三维规划解决方案。通过一项涉及65名真实肿瘤患者的回顾性研究，与经验丰富的临床医生使用现有方法得出的规划结果相比，我们的方法在定性和定量分析方面都取得了良好的效果。这表明我们的方法可以实现一种自动化的术前规划技术，消除主观性，并得出与使用者无关的结果。此外，我们在一个临床病例中展示了我们自动规划流程的潜力，突出了它在临床环境中的适用性。

Method

方法

3.1. Overview

The proposed framework for automatic preoperative planning inmandibular reconstruction with Fibula Free Flap (FFF) is outlined in(Fig. 1). We proposed a neural network approach based on a shapeprior to virtually reconstruct the complete shape of the mandible basedon the remaining healthy structures, as described in Section 3.2. Wedefine 3D function variables within the predicted intact mandibles andestablish a 3D spatial transformation relationship between them andthe path planning, allowing the definition of a mandibular planningscheme from the 3D coordinates passing through the defined variables.(detailed in Section 3.3). The feasible domains of the variables arealso defined in the predicted mandible (detailed in Section 3.4). Thediscrepancy between the morphology of the fibula segments and themandible is quantified using 2D morphometric descriptors, which areexpressed as an objective function. The planning scenario that providesthe best matching shape is determined by optimizing the objectivefunction, as illustrated in Section 3.5.

3.1 概述所提出的用于腓骨游离皮瓣（FFF）下颌骨重建术前自动规划的框架如图1所示。正如3.2节所述，我们提出了一种基于形状先验的神经网络方法，根据剩余的健康结构来虚拟重建下颌骨的完整形状。我们在预测的完整下颌骨内定义三维函数变量，并在这些变量与路径规划之间建立三维空间变换关系，从而能够从经过定义变量的三维坐标来定义下颌骨的规划方案（详见3.3节）。变量的可行域也在预测的下颌骨中进行了定义（详见3.4节）。腓骨节段与下颌骨形态之间的差异使用二维形态测量描述符进行量化，这些描述符被表示为一个目标函数。通过优化该目标函数来确定能提供最佳形状匹配的规划方案，如3.5节所示。

Conclusion

结论

In this study, we present a novel two-stage framework for automated preoperative planning in fibula microvascular graft reconstruction of the mandible. The framework addresses critical challenges inreconstructive surgery by combining deep learning with mathematicaloptimization to achieve both personalized and standardized planning.By encompassing the entire surgical planning process, from morphology prediction to final surgical protocol generation, this comprehensiveapproach ensures a truly automated workflow, significantly reducingthe need for manual intervention.The first stage focuses on reconstructing the pre-disease morphology of the mandible. To achieve this, we employ a deep learningarchitecture that integrates shape prior information based on voxeldata, enabling accurate restoration of the original mandibular shape,even in the presence of significant defects. In cases involving defective or deformed native mandibles, our model predicts the completepre-disease morphology, providing a versatile and adaptable planningreference, especially for defects crossing the midline, where the healthyside cannot serve as a reference. Unlike existing models that discardor oversimplify the shape characteristics of the original mandible,our approach leverages prior knowledge from healthy mandibles torefine predictions. This enables our model to maintain the integrity ofunaffected regions while accurately reconstructing missing segments.Comparative experiments on synthetic data demonstrated that our prediction model significantly outperforms two alternative methods (Abdiet al., 2019; Liang et al., 2020), which tend to neglect essential shapefeatures in their predictions. By incorporating healthy mandible dataduring training, we effectively enhance prediction accuracy.A retrospective study was conducted on 65 patients to evaluate theperformance of the second stage of the framework. The findings ofthis study demonstrate that the proposed approach exhibits superiorperformance in comparison to existing methods (Nakao et al., 2016)and comparable results to those of experienced surgeons across threecritical metrics commonly utilized in fibula mandibular reconstruction:Volume Ratio 𝐸𝑣 , Contour Error 𝐸𝑐 , and Maximum Projection 𝐸𝑝 . Thesemetrics comprehensively reflect the reconstruction quality by assessingvolumetric filling, alignment with the original contour, and the avoidance of unnecessary protrusions. The optimization process embedded inour method ensures accurate and efficient surgical planning, addressingthe clinical requirements for mandibular reconstruction with FFF.Our automated planning method has demonstrated exceptional performance, achieving outcomes comparable to those of experiencedsurgeons through both quantitative and qualitative experiments. Thisaccomplishment represents a significant milestone in user-independentautomated preoperative planning, effectively addressing a critical gapin clinical workflows by eliminating subjectivity and reducing variability in the planning process. Additionally, our method significantlyimproves efficiency, with substantial reductions in planning time anduser interactions (mouse clicks and scrolls). By replicating expert-leveloutcomes while enhancing efficiency, our approach provides a robustsolution to the challenges faced in clinical practice.We also conducted a clinical experiment with a patient diagnosedwith ameloblastoma and a defect crossing the midline. The resultsdemonstrated that our automated planning method successfully addresses the challenges associated with midline-crossing defects, whichare very difficult for complicated manual planning. Notably, this wasachieved without compromising planning reliability while improvingefficiency significantly. These findings highlight the potential of ourmethod to handle a wide range of clinical scenarios, including complexcases that demand precise and consistent outcomes.This study represents a continuation of our previous research (Guoet al., 2022), in which we developed a robotic system aimed at executing manually designed surgical plans. Nevertheless, our current workshifts the focus to automated preoperative planning, exploring its potential to optimize mandibular reconstruction outcomes. This progressfrom manual to automated planning reflects a significant advancement,as the proposed framework leverages a data-driven approach to predictpre-disease morphology and optimize fibula graft placement.Our planning approach is designed for single-barrel fibula flapreconstruction, effectively addressing a broad spectrum of mandibulardefects. While the double-barrel fibula flap technique, recently proposed by Yang et al. (2020), offers advantages in addressing heightdiscrepancies in the anterior mandible, it is limited by the fibula’slength when dealing with larger defects. Nonetheless, we recognize thepotential benefits of this technique and intend to extend our planningalgorithm to support double-barrel reconstruction in the future, therebyenhancing the personalization and applicability of our method. Additionally, current clinical planning considerations are primarily focusedon restoring the mandibular contour, while the inclusion of occlusalrelationships, biomechanical stability, and other factors requires furtherresearch. Specifically, the incorporation of maxillomandibular occlusion into the planning process is currently impractical due to the timingof occlusal restoration and the variability in dental arch defects, particularly in patients with larger defects. The design of surgical guides wasnot included in this phase of the study, as our primary focus was tovalidate the accuracy and efficiency of the automated planning framework itself. In future work, we plan to explore the potential integrationof automated surgical guide design, building upon the geometric datagenerated by our framework to create patient-specific guides that alignwith the surgical plan.The success of our technique in planning mandible reconstructionwith fibula flap indicates its potential applicability not only for this specific procedure but also for other medical applications involving autologous bone grafting. The progress achieved in our retrospective studyprovides strong evidence that our method holds significant promise forfuture clinical applications, assisting surgeons in planning proceduresin a standardized and accurate manner. To facilitate its broad clinicaladoption, we will conduct further experimental evaluations.

在这项研究中，我们提出了一种新颖的两阶段框架，用于在腓骨微血管移植重建下颌骨的术前自动规划。该框架通过将深度学习与数学优化相结合，解决了重建手术中的关键挑战，以实现个性化和标准化的规划。通过涵盖从形态预测到最终手术方案生成的整个手术规划过程，这种全面的方法确保了真正的自动化工作流程，显著减少了对手动干预的需求。第一阶段侧重于重建下颌骨患病前的形态。为了实现这一目标，我们采用了一种深度学习架构，该架构基于体素数据整合了形状先验信息，即使存在明显的缺损，也能够准确恢复原始的下颌骨形状。在涉及到下颌骨存在缺损或变形的病例中，我们的模型可以预测出完整的患病前形态，提供了一个通用且适应性强的规划参考，特别是对于跨越中线的缺损，因为在这种情况下健康的一侧无法作为参考。与现有的一些模型不同，那些模型会丢弃或过度简化原始下颌骨的形状特征，而我们的方法利用了来自健康下颌骨的先验知识来完善预测。这使得我们的模型在准确重建缺失部分的同时，能够保持未受影响区域的完整性。对合成数据进行的对比实验表明，我们的预测模型明显优于另外两种替代方法（阿卜迪等人，2019年；梁等人，2020年），后两种方法在预测中往往会忽略重要的形状特征。通过在训练过程中纳入健康下颌骨的数据，我们有效地提高了预测的准确性。我们对65名患者进行了回顾性研究，以评估该框架第二阶段的性能。这项研究的结果表明，与现有方法（中尾等人，2016年）相比，所提出的方法表现出更优越的性能，并且在腓骨下颌骨重建中常用的三个关键指标（体积比(E_v)、轮廓误差(E_c)和最大投影(E_p)）方面，与经验丰富的外科医生的结果相当。这些指标通过评估体积填充情况、与原始轮廓的对齐程度以及避免不必要的突出，全面地反映了重建质量。我们方法中嵌入的优化过程确保了准确高效的手术规划，满足了腓骨游离皮瓣（FFF）下颌骨重建的临床要求。我们的自动规划方法已展现出卓越的性能，通过定量和定性实验，取得了与经验丰富的外科医生相当的结果。这一成就代表了在与用户无关的自动术前规划方面的一个重要里程碑，通过消除规划过程中的主观性和减少可变性，有效地填补了临床工作流程中的一个关键空白。此外，我们的方法显著提高了效率，大幅减少了规划时间和用户交互（鼠标点击和滚动操作）。通过在提高效率的同时复制专家级的结果，我们的方法为临床实践中面临的挑战提供了一个强有力的解决方案。我们还对一名被诊断患有成釉细胞瘤且存在跨越中线缺损的患者进行了临床实验。结果表明，我们的自动规划方法成功地解决了与跨越中线缺损相关的挑战，而对于复杂的手动规划来说，这类缺损非常棘手。值得注意的是，在显著提高效率的同时，这一结果的取得并没有牺牲规划的可靠性。这些发现突出了我们的方法在处理各种临床场景方面的潜力，包括那些需要精确和一致结果的复杂病例。这项研究是我们之前研究（郭等人，2022年）的延续，在之前的研究中，我们开发了一个旨在执行手动设计的手术方案的机器人系统。然而，我们目前的工作将重点转移到了自动术前规划上，探索其优化下颌骨重建结果的潜力。从手动规划到自动规划的这一进展反映了一个重大的进步，因为所提出的框架利用了一种数据驱动的方法来预测患病前的形态，并优化腓骨移植的位置。我们的规划方法是为单腓骨瓣重建而设计的，有效地解决了广泛的下颌骨缺损问题。虽然杨等人（2020年）最近提出的双腓骨瓣技术在解决下颌前部的高度差异方面具有优势，但在处理较大的缺损时，它受到腓骨长度的限制。尽管如此，我们认识到这种技术的潜在好处，并打算在未来扩展我们的规划算法，以支持双腓骨瓣重建，从而提高我们方法的个性化和适用性。此外，当前的临床规划考虑主要集中在恢复下颌骨的轮廓上，而纳入咬合关系、生物力学稳定性和其他因素则需要进一步研究。具体而言，由于咬合恢复的时机以及牙弓缺损的可变性，特别是在有较大缺损的患者中，目前将上下颌咬合纳入规划过程是不切实际的。在本阶段的研究中没有包括手术导板的设计，因为我们的主要重点是验证自动规划框架本身的准确性和效率。在未来的工作中，我们计划探索自动手术导板设计的潜在整合，利用我们的框架生成的几何数据来创建与手术方案相匹配的针对患者的导板。我们的技术在腓骨瓣下颌骨重建规划方面的成功表明，它不仅对这一特定手术具有潜在的适用性，而且对其他涉及自体骨移植的医疗应用也具有适用性。我们回顾性研究中取得的进展有力地证明，我们的方法对未来的临床应用具有重大的前景，能够帮助外科医生以标准化和准确的方式规划手术。为了促进其在临床上的广泛应用，我们将进行进一步的实验评估。

Figure

图

Fig. 1. Flow diagram of the automated preoperative planning method for mandible reconstruction with fibular free flap. Stage I: The complete shape of the mandible is completedby a neural network that predicts the shape of the defective mandible based on shape prior; Stage II: The difference between the fibula segments and the complete mandible ismeasured in pixels from both transverse and longitudinal views. An objective function is then generated and the planning result is obtained through optimization.

图1：腓骨游离皮瓣下颌骨重建术前自动规划方法的流程图。第一阶段：利用基于形状先验来预测缺损下颌骨形状的神经网络，补全下颌骨的完整形状；第二阶段：从横向和纵向视图，以像素为单位测量腓骨节段与完整下颌骨之间的差异。然后生成一个目标函数，并通过优化得到规划结果。

Fig. 2. Scheme of our mandible shape completion method: After shape warping, the mean mandibular shape is generated. Following registration and PCA, the shape prior *𝑎𝑣𝑔*is obtained. The defective mandible *𝑑𝑒𝑓 𝑒𝑐𝑡* is then registered with the shape prior and input into the neural network to generate the predicted mandibular voxels *𝑝𝑟𝑒* .

图2：我们的下颌骨形状补全方法示意图：在进行形状变形后，生成平均下颌骨形状。经过配准和主成分分析（PCA）后，得到形状先验\(\mathcal{S}_{avg}\)。然后将缺损的下颌骨\(\mathcal{M}_{defect}\)与该形状先验进行配准，并输入到神经网络中，以生成预测的下颌骨体素\(\mathcal{M}_{pre}\) 。

Fig. 3. Scheme of fibula osteotomy plan determined by 𝑖 .

图3：由(\mathcal{P}_{i})确定的腓骨截骨规划示意图。

Fig. 4. Scheme of determination of feasible region

图4：可行区域的确定示意图。

Fig. 5. Predicted mandible segments superimposed on defective mandibles, viewed in3D, axial, sagittal and coronal planes. For differentiation, the defective mandible iswhite and the predicted defect is red

图5：预测的下颌骨节段叠加在缺损下颌骨上的三维视图、轴面视图、矢状面视图和冠状面视图。为了便于区分，缺损的下颌骨为白色，预测的缺损部分为红色。

Fig. 6. Qualitative comparison of mandible reconstruction methods across four cases. Columns correspond to (1) the defective mandible in 𝑡𝑒𝑠𝑡, (2) DCGAN (Abdi et al., 2019),(3) CVAE (Liang et al., 2020), (4) our method without shape priors, and (5) our method with shape priors. The left side of each column shows the predicted reconstruction results(in green), while the right side visualizes the signed distance error between the predicted reconstruction and the ground truth using a colormap

图6：四个病例中下颌骨重建方法的定性比较。各列分别对应（1）测试集(\mathcal{D}_{test})中的缺损下颌骨，（2）深度卷积生成对抗网络（DCGAN，阿卜迪等人，2019）的结果，（3）变分自编码器（CVAE，梁等人，2020）的结果，（4）我们不使用形状先验的方法的结果，以及（5）我们使用形状先验的方法的结果。每一列的左侧展示了预测的重建结果（绿色），而右侧则使用色彩映射图直观呈现了预测的重建结果与真实情况之间的有向距离误差。

Fig. 7. Our automated planning software is implemented using open-source toolkits,enabling the automatic generation of FFF mandible reconstruction planning schemes.The software offers three visualization windows, including a 2D axial view of the skull,a 3D view of the mandible, and a 3D view of the fibula

图7：我们的自动规划软件是使用开源工具包开发的，能够自动生成腓骨游离皮瓣（FFF）下颌骨重建的规划方案。该软件提供了三个可视化窗口，分别是颅骨的二维轴位视图、下颌骨的三维视图以及腓骨的三维视图。

Fig. 8. Qualitative accuracy results for five representative cases, each containing the main process of the planning. The last column shows the manual planning scheme performedby the surgeons.

图8：五个具有代表性病例的定性准确性结果，每个病例都包含了规划的主要过程。最后一列展示了由外科医生执行的手动规划方案。

Fig 9. Statistical comparison of three evaluation metrics for different planning methods (𝑀1 : Proposed method, 𝑀2 : Nakao et al. (2016), 𝑀3 : Manual planning) across twoscenarios (three-segment and two-segment plans): (a) Volume Ratio (𝐸𝑣 ), (b) Contour Error (𝐸𝑐 ), and (c) Maximum Projection (𝐸𝑝 )

图9：针对不同规划方法（M1：所提出的方法，M2：中尾等人（2016年提出的方法），M3：手动规划）在两种情形（三段规划和两段规划方案）下，对三项评估指标的统计比较：(a) 体积比（Ev），(b) 轮廓误差（Ec），以及 (c) 最大投影（Ep）。

Fig. 10. Clinical experiment planning pipeline: Initially, preoperative CT reconstruction was carried out to obtain detailed images of the patients’ mandibles. Following this, themandible was segmented and the range for osteotomy was determined. Subsequently, the comprehensive mandible was automatically completed, and a layer-by-layer similarityassessment was performed to generate the objective function. Finally, the objective function underwent optimization, resulting in the determination of the optimal planning scheme.

图10：临床实验规划流程：首先，进行术前CT重建以获取患者下颌骨的详细图像。在此之后，对下颌骨进行分割并确定截骨范围。接着，自动补全完整的下颌骨，并进行逐层相似度评估以生成目标函数。最后，对目标函数进行优化，从而确定出最佳规划方案。

Fig. 11. (a) Preoperative frontal view. (b) Preoperative craniomaxillofacial 3D model. (c) Orthopantomograph preoperatively. (d) Excision of mandible and lesion. (e) Place theintraoperative osteotomy guide plate. (f) Fibula flaps prepared during the operation. (g) Frontal view 2 weeks postoperatively. (h) Postoperative craniomaxillofacial 3D model. (i)Orthopantomograph postoperatively

图11：(a) 术前正面视图。(b) 术前颅颌面三维模型。(c) 术前曲面断层片。(d) 下颌骨及病变组织的切除。(e) 术中放置截骨导板。(f) 术中制备的腓骨皮瓣。(g) 术后两周的正面视图。(h) 术后颅颌面三维模型。(i) 术后曲面断层片。

Table

表

Table 1Quantitative evaluation results comparing the performance of different predictionmethods using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and RelativeError (RE).

表1 运用迪赛相似系数（DSC）、豪斯多夫距离（HD）以及相对误差（RE），对不同预测方法的性能进行比较的定量评估结果。