Thompson Peter, Khattak Mohammed, Joseph P J, Perry Daniel C, Cootes Timothy F, Lindner Claudia, Karthikappallil Dileep, Zaman Hesham, Airey Grace, Maqsood Saad, Hughes Tom, Ahmad Shuja, McEvoy James, Wilson Graeme, Do Le Ha P, Tariq Fatima, Shah Sohan, Patel Dhawal, McAllister Ross, Singh Dhadwal Anil, Fennelly Joseph, Lloyd William, Varasteh Amir, Almond Kieran, Crouch-Smith Henry
Division of Informatics, Imaging & Data Sciences, The University of Manchester, Manchester, UK.
University of Liverpool, Liverpool, UK.
Bone Joint J. 2025 Jan 1;107-B(1):124-132. doi: 10.1302/0301-620X.107B1.BJJ-2024-0894.
The aims of this study were to develop an automatic system capable of calculating four radiological measurements used in the diagnosis and monitoring of cerebral palsy (CP)-related hip disease, and to demonstrate that these measurements are sufficiently accurate to be used in clinical practice.
We developed a machine-learning system to automatically measure Reimer's migration percentage (RMP), acetabular index (ACI), head shaft angle (HSA), and neck shaft angle (NSA). The system automatically locates points around the femoral head and acetabulum on pelvic radiographs, and uses these to calculate measurements. The system was evaluated on 1,650 pelvic radiographs of children with CP (682 females and 968 males, mean age 8.3 years (SD 4.5)). Each radiograph was manually measured by five clinical experts. Agreement between the manual clinical measurements and the automatic system was assessed by mean absolute deviation (MAD) from the mean manual measurement, type 1 and type 2 intraclass correlation coefficients (ICCs), and a linear mixed-effects model (LMM) for assessing bias.
The MAD scores were 5.7% (SD 8.5%) for RMP, 4.3° (SD 5.4°) for ACI, 5.0° (SD 5.2°) for NSA, and 5.7° (SD 6.1°) for HSA. Overall ICCs quantifying the agreement between the mean manual measurement and the automatic results were 0.91 for RMP, 0.66 for ACI, 0.85 for NSA, and 0.73 for HSA. The LMM showed no statistically significant bias.
The results showed excellent agreement between the manual and automatic measurements for RMP, good agreement for NSA, and moderate agreement for HSA and ACI. The performance of the system is sufficient for application in clinical practice to support the assessment of hip migration based on RMP. The system has the potential to save clinicians time and to improve patient care by enabling more comprehensive, consistent, and reliable monitoring of hip migration in children with CP.
本研究的目的是开发一种能够计算用于脑瘫（CP）相关髋关节疾病诊断和监测的四项放射学测量值的自动系统，并证明这些测量值足够准确，可用于临床实践。
我们开发了一种机器学习系统，以自动测量赖默尔移位百分比（RMP）、髋臼指数（ACI）、头干角（HSA）和颈干角（NSA）。该系统可自动在骨盆X光片上定位股骨头和髋臼周围的点，并利用这些点来计算测量值。该系统在1650例CP患儿的骨盆X光片上进行了评估（682名女性和968名男性，平均年龄8.3岁（标准差4.5））。每张X光片由五名临床专家进行手动测量。通过与平均手动测量值的平均绝对偏差（MAD）、1类和2类组内相关系数（ICC）以及用于评估偏差的线性混合效应模型（LMM），评估手动临床测量与自动系统之间的一致性。
RMP的MAD分数为5.7%（标准差8.5%），ACI为4.3°（标准差5.4°），NSA为5.0°（标准差5.2°），HSA为5.7°（标准差6.1°）。量化平均手动测量值与自动结果之间一致性的总体ICC，RMP为0.91，ACI为0.66，NSA为0.85，HSA为0.73。LMM显示无统计学显著偏差。
结果表明，RMP的手动测量与自动测量之间具有极好的一致性，NSA具有良好的一致性，HSA和ACI具有中等一致性。该系统的性能足以应用于临床实践，以支持基于RMP的髋关节移位评估。该系统有可能节省临床医生的时间，并通过对CP患儿的髋关节移位进行更全面、一致和可靠的监测来改善患者护理。
