目的：

全膝关节置换术(Total knee arthroplasty, TKA)是治疗终末期膝关节疾病的金标准。尽管手术技术和假体设计不断进步，但患者术后的满意度仍未达到理想的“遗忘膝”状态。随着对下肢力线分型研究的深入，越来越多的研究表明术后下肢力线的变化可能与TKA患者的不满意度有关。MacDessi 等人报道的膝关节冠状面对线(Coronal Plane Alignment of the Knee, CPAK)分型受到越来越多国家学者的支持。传统的下肢力线参数测量是一个耗时且主观的过程，且目前关于我国骨关节炎人群的CPAK分型相关研究也较少。鉴于人群中力线的多样性，一些个性化的TKA对线理念被提出。基于CT的个性化截骨导板可以指导术中截骨，且成本较低，在实施个性化对线TKA中有优势。然而，膝关节CT图像的精确分割作为此过程的基础和关键步骤，目前依赖于耗时的手工处理，增加了术前等待时间。人工智能的深度学习能够处理大规模和复杂的医学数据，提高了医学信息的整合和分析效率，但其在下肢力线测量及膝关节CT分割上的应用较少。本研究利用人工智能在图像处理上的优势，开发下肢力线的X线测量及膝关节CT分割的算法，提高TKA术前测量及计划的效率；同时开发基于人工智能的PSI系统，提高个性化对线TKA术中实施的精确度。

方法：

1.基于人工智能的下肢力线测量算法的开发及国人冠状面下肢力线分布的研究

基于2D-Unet卷积神经网络开发了自动测量下肢力线的人工智能算法关键点检测模型，收集479例共计944侧诊断为骨关节炎的膝关节数据集用于训练、测试并验证。由一名有多年临床经验的关节外科医师手工测量测试集中数据的下肢力线参数，并与算法测量的结果进行配对比较。采用散点图描述了骨关节炎人群的CPAK分型分布，并比较了不同性别（男性/女性）、年龄 (＜65岁的中年人，≥65岁的老年人）和体重指数（正常<25kg/m²，肥胖及超重≥25kg/m²)的下肢力线参数及CPAK分型差异。

2.基于人工智能的膝关节CT图像分割算法的开发及验证的研究

网络结构采用3D-Unet卷积神经网络，基于来自318名健康体检患者、膝关节骨关节炎患者、膝关节周围骨折患者的171255张CT图像组成的数据集构建的。使用LOSS学习曲线、Dice系数、交并比(Intersection over union, IOU)，豪斯多夫距离(Hausdorff distance, HD)及平均表面距离(Average surface distance, ASD) 对股骨、胫骨、髌骨和腓骨分割性能进行了评估。同时与Unet，Attention-Unet和2.5D-Unet三种卷积神经网络算法进行了分割性能比较。此外，在正常膝关节、骨关节炎和膝关节骨折患者中进行了分割效果的亚组分析。人工智能分割与手工分割的时间也被比较。

3.基于人工智能的膝关节置换术前规划及个性化截骨导板手术系统的开发及验证的研究

采用3D-Unet和改进的HRNet 神经网络结构开发了基于人工智能的全膝关节置换术的术前规划及个性化截骨导板手术系统。46例计划接受TKA手术的患者同时接受了基于人工智能和胶片模板的假体型号规划，其中22例患者设计并应用了PSI。比较了规划假体型号所需的时间以及人工智能设计PSI和手工PSI设计的时间。按照年龄、性别、体重指数、手术日期匹配了22例行手工TKA的患者作为对照组。比较了基于人工智能和手工模板的假体型号规划的准确性以及PSI组与对照组在术后下肢对线、失血、炎症指标、并发症上的差异。

结果：

1.基于人工智能的下肢力线测量算法的开发及国人冠状面下肢力线分布的研究

关键点检测模型在训练、验证和测试数据集上均表现出了良好的性能，能够准确地预测目标关键点的位置。外科医生和算法在股骨远端外侧角、胫骨近端内侧角、关节线会聚角、髋膝踝角四个测量指标之间无明显偏差，平均测量值误差在0.35度内。且人工智能的测量速度显著快于手工测量（10.09秒/侧vs. 3.48分钟/侧）。在膝关节骨关节炎人群CPAK分型方面，Ⅱ型是最常见的表型（41.3%），其次为V型（26.4%）与I型（16.5%）。不同性别、年龄和体重指数的CPAK分型均以Ⅱ型、V型和I型为主，但在各分型比例和下肢力线参数上差异较大。

2.基于人工智能的膝关节CT图像分割算法的开发及验证的研究

3D-Unet网络结构不仅可以精确分割膝关节骨性结构，而且分割出的膝关节边缘较其他网络平滑，没有出现毛刺或者组织粘连。整体来看3D-Unet网络结构在股骨、胫骨、髌骨和腓骨的分割准确性方面优于其他三种算法，展示了最佳的Dice、IOU、HD和ASD性能。亚组分析中，与其他三种网络结构相比，3D-Unet除在骨关节炎组髌骨的ASD劣于Unet及骨折组中腓骨的HD劣于2.5D-Unet外，在正常膝关节、膝关节骨关节炎和膝关节骨折组中均表现出更高的准确性。在时间上手动分割的膝关节分割时间是3D-Unet时间的433到586倍。

3.基于人工智能的膝关节置换术前规划及个性化截骨导板手术系统的开发及验证的研究

人工智能假体型号规划的时间与胶片模板术前规划时间无差异(p＞0.05)，基于人工智能设计PSI所需时间显著快于手工设计时间(p <0.05)，从CT处理到PSI打印完成的时间为19.78±2.36小时。基于人工智能的股骨及胫骨假体型号规划准确度均优于胶片模板规划。此外，PSI提高了下肢整体力线的准确度并降低了术后失血量(p <0.05)。两组患者在炎症指标、住院天数、并发症上无统计学差异(p＞0.05)。

结论：

基于人工智能的下肢力线测量算法提高了下肢力线参数测量的准确度及速度；我国骨关节炎人群的CPAK分型主要以Ⅱ型、V型和I型为主，不同性别、年龄和体重指数的CPAK分型基本一致，但在各下肢力线参数上差异较大。更好地了解膝关节表型的分布及变异性为实施个性化的TKA提供了理论基础。

基于3D-Unet的膝关节CT自动分割人工智能算法可以精确地分割股骨、胫骨、髌骨、腓骨，并且在多种膝关节疾病的应用场景中验证了其稳定性、准确性及快速性。这种便捷、稳定的膝关节 CT 图像分割网络有望提高个性化截骨导板、导航和机器人辅助的全膝关节置换术的术前规划效率。

基于人工智能的全膝关节置换术的术前规划及个性化截骨导板手术系统可显著减少PSI 设计所需的时间，同时不增加假体规划的时间；能准确预测 TKA 患者的假体型号，提高下肢对线的整体准确性并可以减少术后失血，是人工智能在骨科应用的重要补充。

Objective:

Total knee arthroplasty (TKA) is the gold standard for treating end-stage knee joint disease. Despite continuous advancements in surgical techniques and prosthesis design, patient satisfaction has yet to reach the desired state of "forgotten knee." With the deepening research on lower limb alignment classification, an increasing number of studies suggest that changes in postoperative lower limb alignment may be associated with dissatisfaction among TKA patients. The Coronal Plane Alignment of the Knee (CPAK) classification proposed by MacDessi et al. provides a convenient and comprehensive system for describing alignment, garnering increasing support from scholars worldwide. Traditional measurement of lower limb alignment parameters is a time-consuming and subjective process, and there is a lack of research on CPAK classification in the osteoarthritis population in our country. Based on the personalized alignment distribution within populations, various concepts aimed at improving postoperative function and satisfaction of patients have gradually been proposed. Patient-specific instrumentation (PSI) based on computed tomography (CT) imaging can facilitate intraoperative bone resection at a lower cost, providing advantages in implementing personalized alignment TKA. However, the precise segmentation of knee joint CT images serves as the foundation and crucial step in this process. Currently, manual processing methods for CT imaging are time-consuming, increasing preoperative waiting time. Deep learning in artificial intelligence has the capability to handle large-scale and complex medical data, thereby improving the integration and analysis efficiency of medical information, but its application in lower limb alignment measurement and knee joint CT segmentation is relatively limited. This study leverages the advantages of artificial intelligence in image recognition to develop algorithms for X-ray measurement of lower limb alignment and segmentation of knee joint CT images, aiming to enhance the efficiency of preoperative measurement and planning for TKA. Simultaneously, we develop a patient-specific instrumentation system based on artificial intelligence to improve the accuracy of implementation in personalized alignment TKA.

Methods:

Development of an artificial intelligence-based alignment measurement method and research on alignment classification in Chinese population.

An artificial intelligence method for automatic measurement of lower limb alignment was developed using a 2D-Unet convolutional neural network. A total of 944 knees from 479 individuals diagnosed with osteoarthritis were utilized for training, testing, and validation. Lower limb alignment parameters of the testing set were manually measured by an orthopedic surgeon with years of clinical experience and compared with the method's measurements. The distribution of CPAK classifications in the osteoarthritis population was depicted using scatter plots, and differences in lower limb alignment parameters and CPAK classifications among different genders (male/female), ages (<65 years old, ≥65 years old), and body mass index (normal <25 kg/m², overweight and obese ≥25 kg/m²) were compared.

Development and validation of an artificial intelligence-based knee joint CT segmentation method.

A 3D-Unet convolutional neural network was adopted for network structure, based on a dataset comprising 171,255 CT images from 318 healthy individuals, knee osteoarthritis patients, and patients with knee fractures. Evaluation of segmentation performance for the femur, tibia, patella, and fibula was conducted using LOSS learning curves, Dice similarity coefficient (Dice), Intersection over union (IOU), Hausdorff distance (HD), and Average surface distance (ASD). Performance comparison of segmentation was carried out with three other convolutional neural network methods: Unet, Attention-Unet, and 2.5D-Unet. Subgroup analyses were performed for normal knee joints, knee osteoarthritis, and knee fracture images. The time for artificial intelligence segmentation and manual segmentation was also compared.

Development and validation of an artificial intelligence-based preoperative planning and patient-specific instrumentation system for total knee arthroplasty.

An artificial intelligence-based preoperative planning and patient-specific instrumentation system for TKA was developed using 3D-Unet and an improved HRNet neural network structure. Forty-six patients scheduled for TKA surgery received both artificial intelligence-based and acetate template-based prosthesis planning simultaneously, with 22 patients undergoing PSI design and application during surgery. The time required for planning prosthesis size and the time required for designing PSI using artificial intelligence were compared with those of manual PSI design. Twenty patients undergoing manual TKA were matched as controls based on age, gender, body mass index, and surgery date. The accuracy of size planning based on artificial intelligence and manual templates, as well as differences in lower limb alignment, blood loss, inflammatory indicators, and complications between the PSI group and the control group, were compared.

Results:

Development of an artificial intelligence-based alignment measurement method and research on alignment classification in Chinese population.

The key point detection model demonstrated good performance on the training, validation, and testing datasets, accurately predicting the positions of target key points. There was no significant deviation between the four measurement indicators between surgeons and the method, with an average measurement error within 0.35 degrees. Regarding CPAK classifications, Type II was the most common phenotype (41.3%), followed by Type V (26.4%) and Type I (16.5%). CPAK classifications among different genders, ages, and body mass index categories predominantly consisted of Types II, V, and I, but there were substantial differences in the proportions and lower limb alignment parameters among the classifications.

Development and validation of an artificial intelligence-based knee joint CT segmentation method.

The 3D-Unet network structure not only accurately separated the bony structures of the knee joint but also produced smoother knee joint edges compared to other networks, without spurs or tissue adhesions. Overall, the 3D-Unet network structure demonstrated superior accuracy in segmentation of the femur, tibia, patella, and fibula compared to the other three networks, exhibiting the best performance in terms of Dice, IOU, HD, and ASD. In subgroup analysis, compared with the other three network structures, 3D-Unet showed higher accuracy in normal knee joints, knee osteoarthritis, and knee fracture groups, except for the ASD of the patella in the osteoarthritis group, which was inferior to Unet, and the HD of the fibula in the fracture group, which was inferior to 2.5D-Unet. In terms of time, the manual segmentation time for knee joint segmentation was approximately 433 to 586 times that of 3D-Unet.

Development and validation of an artificial intelligence-based preoperative planning and patient-specific instrumentation system for total knee arthroplasty.

There was no difference in the time required for artificial intelligence-based prosthesis planning compared to acetate template-based planning (p > 0.05), while the time required for artificial intelligence-based PSI design was significantly faster than manual design time (p < 0.05), with a time from CT processing to completion of PSI printing of 19.78± 2.36 hours. The accuracy of femoral and tibial prosthesis sizes based on artificial intelligence was superior to that of acetate template planning. Furthermore, PSI significantly improved the overall accuracy of lower limb alignment and reduced postoperative blood loss (p < 0.05). There were no statistically significant differences between the two groups in terms of inflammatory indicators, length of hospital stay, and complications (p > 0.05).

Conclusion:

The AI-based lower limb alignment measurement method enhances the accuracy and efficiency of lower limb alignment parameter measurement. The distribution of CPAK classifications in the Chinese osteoarthritis population is primarily composed of Types II, V, and I, with generally consistent patterns among different genders, ages, and body mass index categories, although significant differences exist in lower limb alignment parameters. A better understanding of population variations in knee phenotypes can assist orthopedic surgeons in formulating more personalized alignment strategies.

The knee joint CT automatic segmentation artificial intelligence method based on 3D-Unet can accurately segment the femur, tibia, patella, and fibula, demonstrating stability, accuracy, and rapidity in various applications of knee joint diseases. This convenient and stable knee joint CT image segmentation network is expected to improve the efficiency of preoperative planning for PSI, navigation, and robotic-assisted total knee arthroplasty.

The artificial intelligence-based preoperative planning and patient-specific instrumentation system for TKA significantly reduces the time required for PSI design without increasing the time for size planning, accurately predicts prosthesis sizes for TKA patients, improves the overall accuracy of lower limb alignment, and reduces postoperative blood loss. It serves as an important supplement to the application of artificial intelligence in orthopedics.