【中图分类号】 R681【文献标识码】 A DOI：10.12114/j.issn.1007-9572.2023.0919

【Abstract】BackgroundAs China's aging population continues to grow，the incidence of osteoporosis has been steadilyincreasing，posingasignificanthealthchallngefortheelderlypopulation.Furthermore，thehighcostofdiagnosingand treating osteoporosishighlightstheimportanceofearlydiagnosisasakeystrategytoreducebothpatientsuferingandhealthcare expenses.ObjectiveTheobjectiveofthisstudyis todevelopachestandabdominalbone mineraldensity（BMD）measurement model using conventionalchest and abdominal CTscans，with deep neural networksand machinelearning algorithms.The abdominalBMDmodelissubsequentlyemployedtocalibrate thechest BMDmeasurements，withthegoalofenablingautomated BMD measurement and thediagnosis of osteoporosis.MethodsThis retrospective studycollected 702 patients from Suining Central Hospital in Sichuan Province whounderwentboth chest CTscansandquantitativeCT（QCT）examinations during the periodfrom March 2O2 to June 2023（spaning appoximatelyoneyear）asresearch subjects.Among them，532 patients were randomly divided into a training set（426 cases， 80% ）and a validation set（106 cases， 20% ）.An additional 170 patients were included inthe internal testing set.Thisstudyused thediagnosticresultsof QCTas thereferencestandardand employs machinelearning methods such aslogisticregresion，stochastic gradient descent，andrandom forest toconstructosteoporosis classificationmodelsandbone densityregression models forthechestandabdomen，the established model was also tested internally. The performance of the classification models was evaluated using sensitivity，specificity，accuracy，precision，and areaunderthereceiver operating characteristiccurve（AUC），while regresion model performance was assessed using mean absolute error （ MAE ），root mean square error （ RMSE ），and R -squared. ResultsThe results showed that the AUC values forthe osteoporosisclasificationmodels inthevalidationsetwereO.948forthechest modelandO.968fortheabdominalmodel. The meanabsoluteerorsof the BMDregression modelswere10.534and9.449，respectively.Inthe internaltestingset，the AUC valuesfortheclasificationmodelswere0.905and0.926，andtheMAEfortheregresionmodelswere9.255and7.924， respectively.Aftercalibration，theAUCand MAEof thechest BMD measurementmodelinthevalidationset improved to 0.967 and10.511，respectively.ConclusionTheAI-basedchestandabdominal BMD measurementsdemonstrateahighcorrelation andconsistency with QCT measurements，efectively diagnosing osteoporosis.Thecalibrated chest BMD measurement model furtherenhancesdiagnosticperformanceandoferssignificantpotentialfortheapplicationofchest CTsans inopportunistic osteoporosis screening.

【Key words】Osteoporosis；Bone mineral density；CT plain scan； Deep learning；Machine learning

骨质疏松是一种慢性骨骼疾病，其特点是低骨密度（bonemineraldensity，BMD）和骨组织微结构恶化［1］。随着我国老龄化趋势的逐步加深，骨质疏松的发病率逐渐攀升，且国内外诊断及治疗骨质疏松的成本较为高昂，甚至有超过一半的骨质疏松患者从未接受过骨质疏松筛查［2]。因此骨质疏松的早期诊断成了降低患者疾病痛苦及治疗成本的关键。目前，骨质疏松的早期诊断主要依靠骨密度的测定，常用的骨密度测量方法包括定量超声系统（quantitativeultrasound system，QUS）、定量计算机断层扫描（quantitativecomputedtomography，QCT）和双能X射线骨密度仪（dualenergyX-ray absorptiometry，DXA）[3]。但由于QUS存在检查准确性能低，QCT需要模具及复杂的后处理，DXA忽略了骨骼的形态和结构并且人均设备量不足等缺点，国内外许多学者提出通过人工智能在CT平扫图像上测量骨密度的方法克服上述检查手段存在的局限性，但建立出来的模型通常只进行了简单的验证校准，模型诊断性能无法得到保证。根据《2018版中国定量CT（QCT）骨质疏松症诊断指南》［4]，由于QCT测量的是真正的体积骨密度，能够更敏感地反应骨质疏松的骨密度变化，而且不受脊柱增生退变和血管钙化等因素的影响，因此本研究以QCT的测量结果作为“金标准”，基于CT平扫图像通过人工智能技术分别建立胸部和腹部骨密度测定模型，并通过腹部测定模型对胸部的骨密度测定进行校准，期望能为提升模型诊断性能的同时为胸部CT平扫在骨质疏松症的机会性筛查中的应用发展提供潜力，更好促进人工智能技术在骨质疏松诊断方面的发展，并为其他疾病的诊断模型建立提供新的思路。……