深度学习在全长腿X线片上评估术前术后下肢冠状面力线的研究
该研究基于2419张标注全长腿X线片开发两阶段深度学习算法,检测关节并标记膝关节手术状态,再定位关键点计算髋膝踝角(HKAA)、股骨远端外侧角(LDFA)和胫骨近端内侧角(MPTA)。在239张测试集上,HKAA平均绝对误差(MAE)为0.38°,LDFA为1.45°,MPTA为1.17°,且术前膝关节表现略优于术后。
要点
- 算法可自动测量HKAA、LDFA和MPTA,不受关节植入物状态影响。
- HKAA的MAE为0.38°,95%测量与手动测量相差在1°以内。
- 术前膝关节各角度测量表现略优于术后膝关节。
核心结论
- 深度学习算法能准确测量全长腿X线片上的HKAA、LDFA和MPTA,不受关节植入物状态影响。
- 高准确度和低MAE(尤其HKAA)提示该工具可用于术前评估或术后随访的力线评估。
研究设计与样本
- 从机构登记库中采样2419张标注全长腿X线片,包含术前和术后的膝、髋、踝关节混合图像。
- 由三名标注者手动放置边界框和关键点,训练两阶段深度学习算法:第一阶段检测关节并标记膝关节手术状态,第二阶段用四个关键点模型定位计算角度所需关键点。
- 在239张独立测试集上评估,以平均绝对误差(MAE)为主要结局指标。
主要发现
- 所有测量的总体MAE:HKAA为0.38°,LDFA为1.45°,MPTA为1.17°。
- HKAA有95%的测量值与手动测量相差在1°以内。
- 与手动测量相差在3°以内的测量比例:HKAA为91%,LDFA为95%,MPTA为99%。
- 术前膝关节在所有角度上的表现略高于术后膝关节。
关键数值与推荐
- HKAA的MAE为0.38°,95%测量在手动测量1°以内。
- LDFA的MAE为1.45°,MPTA的MAE为1.17°。
- 相差3°以内的比例:HKAA 91%、LDFA 95%、MPTA 99%。
临床怎么用
- 该工具可在膝关节置换术前规划和术后随访中自动测量下肢冠状面力线参数。
- 尤其HKAA准确度高,适合需要精确力线评估的临床场景。
局限性
- 术后图像的泛化性和性能可能通过进一步训练提升。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Coronal plane lower-limb alignment, defined by metrics including the hip-knee-ankle angle (HKAA), lateral distal femoral angle (LDFA), and medial proximal tibial angle (MPTA), is useful for surgical planning and postoperative assessment in knee arthroplasty. The purpose of this study was to develop and validate a flexible deep learning algorithm capable of accurately measuring HKAA, LDFA, and MPTA on long-leg radiographs, independent of the operative status of any joint in the image. METHODS: We sampled 2,419 annotated long-leg radiographs from our institutional registry, which included a mixture of pre- and postoperative knees, hips, and ankles. There were three annotators who manually placed bounding boxes and keypoints to train a two-stage deep learning algorithm. The first stage detects joints and labels knee operative status, and the second stage uses four key point models to locate the necessary key points for angle calculation. Performance was evaluated on a held-out test set of 239 images with mean absolute error (MAE) as the primary outcome metric. RESULTS: The end-to-end algorithm demonstrated high accuracy comparable to manual measurements. The overall MAE for all measurements was 0.38° for HKAA, 1.45° for LDFA, and 1.17° for MPTA. For HKAA, 95% of measurements were within 1° of the manual measurement. The proportion of measurements within 3° of the manual measurement was 91, 95, and 99% for HKAA, LDFA, and MPTA, respectively. Performance was slightly higher for preoperative knees compared to postoperative knees across all angles. CONCLUSIONS: We developed a flexible deep learning algorithm to accurately measure HKAA, LDFA, and MPTA on long-leg radiographs irrespective of the implant status of the joints. The high accuracy and low MAE, particularly for HKAA, suggest this tool can perform alignment assessments during the preoperative workup or postoperative follow-up of patients. Further training may improve the generalizability and performance on postoperative images.
引用
- Mulford KL, Khela M, Wang E, et al. The Assessment of Coronal Plane Lower-Limb Alignment on Pre- and Postoperative Long-Leg Radiographs Using Deep Learning. J Arthroplasty. 2026. PMID: 41999795. doi:10.1016/j.arth.2026.04.036
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。