Park Nancy, Sieberer Johannes, Manafzadeh Armita, Hackbarth Rieke-Marie, Desroches Shelby, Ghankot Rithvik, Lynch John, Segal Neil A, Stefanik Joshua, Felson David, Fulkerson John P
Yale University, New Haven, Connecticut, U.S.A.
University Medical Center Hamburg-Eppendorf, Hamburg, Germany.
Arthrosc Sports Med Rehabil. 2024 Nov 5;7(2):101036. doi: 10.1016/j.asmr.2024.101036. eCollection 2025 Apr.
To assess the inter- and intrarater reliability of 21 anatomical landmarks initially placed with an artificial intelligence algorithm and then manually verified with human input.
Thirty computed tomography scans of the knees of participants from the Multicenter Osteoarthritis Study (MOST) ages 45 to 55 years were included. Approximately one-half experienced progression of patellofemoral osteoarthritis, defined as an increased cartilage score in the patellofemoral compartment on magnetic resonance imaging over 2 years. The algorithm automatically placed 19 anatomic landmarks on the femur, tibia, and patella. An additional 2 landmarks were added manually. Two landmark reviewers separately reviewed all 30 scans and verified all landmarks. After 2 weeks, one reviewer repeated the process for the same dataset. The mean Euclidean distance of manual landmark displacement, mean absolute disagreement between and within raters, and intraclass correlation coefficients for inter- and intrarater reliability were calculated.
All landmarks had excellent inter-rater reliability. The tibial and femoral shaft centers had intraclass correlation coefficients (ICCs) of 1, indicating their positions did not differ. Seventeen landmarks had ICCs between 0.90 and 0.99 and the tibial tuberosity had an ICC of 0.87. Intrarater reliability scores were 1 for 16 landmarks and between 0.90 and 0.99 for the remaining 5.
There was excellent agreement on the locations of all 21 landmarks evaluated in this study.
The potential role of artificial intelligence in medical imaging and orthopaedic research is a growing area of interest. The excellent reliability demonstrated across multiple landmarks in our study reveals the potential for semiautomated 3-dimensional methods to enhance precision of anatomical measurements of the knee over 2-dimensional methods.
评估最初由人工智能算法放置、随后经人工输入进行手动验证的21个解剖标志点的评分者间和评分者内信度。
纳入多中心骨关节炎研究（MOST）中30例年龄在45至55岁参与者的膝关节计算机断层扫描图像。约一半参与者出现髌股关节炎进展，定义为在2年期间磁共振成像显示髌股关节软骨评分增加。该算法自动在股骨、胫骨和髌骨上放置19个解剖标志点。另外手动添加2个标志点。两名标志点审阅者分别审阅所有30幅扫描图像并验证所有标志点。2周后，一名审阅者对同一数据集重复该过程。计算手动标志点位移的平均欧几里得距离、评分者间和评分者内的平均绝对差异以及评分者间和评分者内信度的组内相关系数。
所有标志点均具有出色的评分者间信度。胫骨干和股骨干中心的组内相关系数（ICC）为1，表明其位置无差异。17个标志点的ICC在0.90至0.99之间，胫骨结节的ICC为0.87。评分者内信度得分中，16个标志点为1，其余5个标志点在0.90至0.99之间。
本研究中评估的所有21个标志点的位置具有高度一致性。
人工智能在医学成像和骨科研究中的潜在作用是一个日益受关注的领域。我们研究中多个标志点所显示的出色信度揭示了半自动三维方法相较于二维方法在提高膝关节解剖测量精度方面的潜力。
