Research On Quantitative Prediction Methods Of Bone Strength Based On Clinical QCT Images

Osteoporosis and degenerative change are common conditions among aging adults,and they can cause bone mass reduction,bone microstructural degeneration,and changes in morphology and density distribution.The major complication of the above conditions is pathological fracture,which can decrease patients’mobility with a substantial impact on life quality,and even significantly increase morbidity and mortality.The causation of pathological fracture is the reduced bone strength.Therefore,it is necessary to accurately quantify bone strength to help clinicians assess fracture risk and develop individualized therapeutic interventions.Accordingly,the aims of this dissertation were:（1）To explore the effects of image resolution and element size on the quantitative computed tomography-based finite element analysis（QCT/FEA）outcomes,and to provide a theoretical basis for improving the accuracy of QCT/FEA-calculated bone strength.（2）To develop strength prediction models for the vertebral body,proximal femur,and degenerated vertebra based on QCT images by using machine learning and multivariable linear regression（MLR）,and to provide new methods for clinical bone strength prediction and fracture risk assessment.This dissertation was composed of the following four parts:（1）The effects of image resolution and element size on the QCT/FEA outcomes were explored.Finite element（FE）models of bovine vertebral body were developed based on the QCT images with two resolutions（0.6mm and 1mm slice thicknesses）,and then they were meshed by hexahedral elements with different sizes,namely,0.41×0.41×0.41mm³,0.41×0.41×0.6mm³,0.41×0.41×1mm³,1×1×1mm³,2×2×2mm³,and 3×3×3mm³.Elastic moduli and yield strengths derived from the above FE models were compared.The mechanical parameters of the FE models（meshed using hexahedral elements with the same element size）developed from QCT images with0.6mm and 1mm slice thicknesses were almost the same.In particular,the computational accuracy of the FE models developed from QCT images with 1mm slice thickness was satisfactory,and their computational cost was relative low.Elastic moduli of the FE models（developed based on the same resolution images）with the element sizes larger than 0.41×0.41×1mm³ were significantly different from those of the FE models with the element sizes less than 0.41×0.41×1mm³.In conclusion,it was recommended that FE models with the element size of 0.41×0.41×0.6mm³developed from the QCT images with 1mm slice thickness could be utilized to predict the mechanical parameters of vertebral body.This study provided a theoretical basis for the selection of image resolution and element size,which may improve the accuracy of bone strength prediction.（2）A noninvasive method of vertebral strength prediction was provided based on clinical QCT images by using machine learning.Eighty elderly men with QCT data of lumbar spine were randomly selected from the Mr OS cohort.A total of 58 parameters of L1 vertebral body were extracted from QCT images of each subject,including grayscale distribution（15 parameters）,grayscale values of partitioned regions（39parameters）,bone mineral density（BMD）,structural rigidity,axial rigidity,and the product of BMD and the minimum cross-sectional area of vertebral body(BMD×A_min).Feature selection and dimensionality reduction were used to simplify the 58 parameters.L1 vertebral strengths were calculated by QCT/FEA.Artificial neural network and support vector regression（SVR）models were developed to predict vertebral strength.It was shown that high accuracy was achieved by using five features(grayscale value of the 60%percentile,grayscale values of three specific partitioned regions,and BMD×A_min)or nine principal components to predict vertebral strength.This study proposed an effective method of vertebral strength prediction,which has great potential in the noninvasive assessment of vertebral fracture.（3）A noninvasive method of proximal femoral strength prediction was provided based on clinical QCT images by using machine learning.Eighty elderly men with QCT data of hip region were randomly selected from the Mr OS cohort.A total of 50parameters of right proximal femur were extracted from QCT images of each subject,including grayscale distribution（15 parameters）,regional cortical bone mapping（CBM）measurements（30 parameters）,and three-dimensional geometric parameters（5 parameters）.Feature selection and dimensionality reduction were used to simplify the 50 parameters.Proximal femoral strengths were calculated by QCT/FEA,and SVR models were developed to predict proximal femoral strength.For feature selection,the best prediction performance of SVR models was achieved by integrating the grayscale value of 30%percentile and specific regional CBM measurements;and for dimensionality reduction,the best prediction performance of SVR models was achieved by extracting principal components with eigenvalues greater than 1.0.The femoral strengths predicted from the well-trained SVR models were in good agreement with those derived from QCT/FEA.This computer-aided diagnostic tool shows great potential in clinical applications,and can be utilized in routine bone health assessments.In addition,this study may improve the ability to noninvasively evaluate fracture risk and develop individually targeted interventions to reduce bone fractures.（4）Degeneration-related variations of morphology and density distribution in vertebrae were investigated,and the effects of above variations on bone strength were evaluated.The statistical shape model（SSM）and statistical appearance model（SAM）were developed based on the QCT images of L1 vertebrae from 75 elderly men,and variations in shape and density distribution of degenerated vertebrae（mild,moderate,and severe grades）were quantified with principal component（PC）modes.Moreover,QCT/FEA was used to calculate compressive strength for each L1 vertebra,the associations between compressive strength and PC modes were evaluated by MLR.There were significant differences in the five specific PC modes（SSM 5,SSM 11,SAM 2,SAM 3,and SAM 5:mainly identifying height and width of vertebral body,the density variations in posterior element,vertebral body,posterior region of spinal foramen,cortical bone of vertebral body,and spinous process）among the vertebrae with different degenerative states,which could be used to discriminate degeneration grade.A reasonable accuracy of bone strength prediction was achieved by using the four PC modes（SSM 1,SAM 1,SAM 4,and SAM 5:mainly identifying the geometrical scaling,the density variations in vertebra,cortical bone of vertebral body,and spinous process）to construct the MLR model.These findings have implications for assisting clinicians in pathological diagnosis,fracture risk assessment,implant selection,and preoperative planning.Furthermore,this study may help to improve understanding of the relevant mechanisms of lumbar degeneration.