基于腹部和盆腔CT的机器学习机会性筛查骨质疏松和骨量减少
本研究回顾性纳入394例50岁以上同时有腹部/盆腔CT和DXA的患者,对L1-L4、骨盆和骶骨进行分割获取平均CT衰减值,比较L1单骨阈值、单变量逻辑回归、LASSO和SVM模型预测骨质疏松/骨量减少的性能。SVM模型(准确率0.892,AUC 0.886)优于L1阈值173.9 HU(准确率0.725,AUC 0.739)、单变量逻辑回归(准确率0.767,AUC 0.533)和LASSO模型(准确率0.817,AUC 0.711)。
要点
- 多骨CT衰减值联合临床数据的SVM模型预测骨质疏松/骨量减少优于单用L1衰减值。
- L1单骨CT衰减阈值173.9 HU可用于机会性筛查,但性能低于多变量模型。
- 腹部/盆腔CT可机会性评估骨密度,提示报告时可关注骨质疏松风险。
核心结论
- 使用腰椎和骨盆多个骨骼的CT衰减值及临床数据的SVM模型预测骨质疏松/骨量减少的能力优于单用L1 CT衰减值或LASSO模型。
研究设计与样本
- 回顾性研究,纳入394例50岁及以上患者,均有腹部和盆腔CT及6个月内的DXA检查。
- 对L1-L4椎体、骨盆和骶骨进行容积分割,获取各骨平均CT衰减值。
- 数据随机分为训练/验证集(n=274,70%)和测试集(n=120,30%)。
- 比较L1 CT衰减阈值、单变量逻辑回归、LASSO和径向基核SVM模型的预测性能,使用DeLong检验。
诊断标准与关键数值
- SVM模型在测试集的准确率为0.892,AUC为0.886。
- L1 CT衰减阈值173.9 HU的准确率为0.725,AUC为0.739(与SVM比较p=0.010)。
- 单变量逻辑回归模型的准确率为0.767,AUC为0.533(与SVM比较p<0.001)。
- LASSO模型的准确率为0.817,AUC为0.711(与SVM比较p=0.007)。
- 所有CT衰减测量值均对骨质疏松/骨量减少有预测作用(p<0.001)。
临床应用与阅片要点
- 在腹部/盆腔CT报告中,可关注腰椎和骨盆骨的CT衰减值,提示骨质疏松/骨量减少风险。
- L1 CT衰减值低于173.9 HU可作为机会性筛查的参考阈值,但需注意其预测性能低于多变量模型。
局限性
- 本研究为回顾性设计,样本量有限(394例),且仅纳入50岁及以上人群。
- 未报告外部验证,模型泛化能力需进一步研究。
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)。
英文原文(摘要)· 点击展开
OBJECTIVES: To use multivariable machine learning using the computed tomography (CT) attenuation of each of the bones in the lumbar spine, pelvis, and sacrum, to predict osteoporosis/osteopenia.
METHODS: This was a retrospective study of 394 patients aged 50 years or older with CT scans of the abdomen and pelvis and dual-energy x-ray absorptiometry (DXA) scans obtained within 6 months of each other. Volumetric segmentations were performed for each of the bones from L1-L4 vertebrae, pelvis, and sacrum to obtain the mean CT attenuation of each bone. The data was randomly split into training/validation (n = 274, 70%) and test (n = 120, 30%) datasets. The CT attenuation of the L1 vertebrae, univariate logistic regression, least absolute shrinkage and selection operator (LASSO), and support vector machines (SVM) with radial basis function (RBF) were used to predict osteoporosis/osteopenia. The performance of using the CT attenuation at L1 to the univariate logistic regression, LASSO, and SVM models were compared using DeLong's test in the test dataset.
RESULTS: All CT attenuation measurements were predictive of osteoporosis/osteopenia (p < 0.001 for all). The SVM model (accuracy = 0.892, AUC = 0.886) outperformed the models using the CT attenuation of threshold of 173.9 Hounsfield units (HU) at L1 (accuracy = 0.725, AUC = 0.739, p = 0.010), the univariate logistic regression model (accuracy = 0.767, AUC = 0.533, p < 0.001) and the LASSO model (accuracy = 0.817, AUC = 0.711, p = 0.007) to predict osteoporosis/osteopenia.
CONCLUSION: A SVM model using the CT attenuations of multiple bones within the lumbar spine and pelvis and clinical data has a better ability to predict osteoporosis/osteopenia than using the CT attenuation of L1 or a LASSO model.
KEY POINTS: • Multivariable SVM model using the CT attenuation of multiple bones and clinical/demographic data was more predictive than using the CT attenuation at L1 only.
引用
- Sebro, De la Garza-Ramos. Opportunistic screening for osteoporosis and osteopenia from CT scans of the abdomen and pelvis using machine learning. Eur Radiol. 2022. PMID: 36166085. doi:10.1007/s00330-022-09136-0
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。