Classification Of Osteoporosis Diagnosis From X-ray Plain Film Of Femur Neck Using Computer Aided Diagnosis

Purpose:To explore the application value of computer-aided diagnosis in the diagnosis of osteoporosis.Firstly,the texture features of Ward triangle of femoral neck were extracted by texture feature analysis.BP neural network classifier and SVM classifier were used to identify osteoporosis.Methods:The original data were collected and grouped.The pelvic X-ray images were obtained from the second affiliated Hospital of Dalian Medical University.89 cases of pelvic X-ray examination were selected.Bone mineral density(BMD)was measured by dual-energy X-ray absorptiometry(DXA)in all patients within 6 months.According to the results of BMD,the patients were divided into three groups S0(30 cases of normal BMD),S1(28 cases of osteopenia)and S2(31 cases of osteoporosis.).Computer aided Diagnostic Analysis: first step,The region of interest is 20 × 20.Each patient extracts 4-8 ROIs.A total of 529 ROIs are obtained,of which 182 in S0 group,178 in S1 group,169 in S2 group.The texture feature parameters of each ROI are extracted.Firstly,20 texture feature parameters of each ROI are extracted by gray co-occurrence matrix method in 4 directions,such as correlation,dark clustering,entropy and so on,and the average value of each parameter is obtained.Then four parameters of each ROI are extracted: run length inhomogeneity factor,gray inhomogeneity factor,long run factor and short run factor,and 24 texture feature parameters are obtained.Feature screening: drawing box-like images according to extracted texture features to screen features,4th steps,using BP neural network classifier and SVM classifier to identify S0-S 2 3 groups of ROI and draw ROC curve.The accuracy of BP neural network classifier and SVM classifier is compared.Results:1.The result of texture feature selection: according to the extracted texture features,we draw 24 box images corresponding to the quantity of texture features,and delete 11 feature parameters with more overlap between the two groups.Thirteen texture feature parameters(autocorrelation coefficient,entropy,contrast,homogeneity,maximum probability,variance,mean,sum entropy,variance,difference variance,difference entropy)were obtained.Run length unevenness factor and gray unevenness factor.2.The result of BP neural network classifier :Classifying accuracy rates of S0-S1 was 71.24%,under the corresponding ROC curve was 0.6921,Classifying accuracy rates of S0-S2 was 72.36%,under the corresponding ROC curve was 0.7011.Classifying accuracy rates of S1-S2 was 61.56%,the area under the corresponding ROC curve was 0.6348.3.The result of SVM classifier: Classifying accuracy rates of S0-S1 was 73.52%,under the corresponding ROC curve was 0.7377.Classifying accuracy rates of S0-S2 was 76.34%,under the corresponding ROC curve is0.7844.Classifying accuracy rates of S1-S2 was 64.55%,the area under the corresponding ROC curve is 0.6758.Conclusion:The SVM and BP neural network classifiers based on texture feature can be used to identify the osteoporosis plain film image.The recognition rate of SVM classifier is slightly higher than that of BP neural network classifier.