Preliminary Study On Classification Of Osteoporosis Diagnosis From CT Of Lumbar Spine Using Computer Aided Diagnosis

Purpose: Texture characteristics of trabeculae bone in lumbar CT images were analyzed.Then,back-propagation(BP)neural network classifier,regression tree classifier and support vector machine(SVM)classifier were used to classify osteoporosis.To explore the feasibility of computer-aided diagnosis based on lumbar CT in the diagnosis of osteoporosis.Methods:(1)Image acquisition and grouping: 139 patients who received dual-energy X-ray absorptiometry(DXA)examination and CT analysis of the lumbar spine within three months in the second affiliated hospital of Dalian medical university were retrospectively reviewed.All the patients were divided into three groups according to the T-scores of DXA results: S0 group(44 cases with normal bone mineral density),S1 group(47 cases with osteopenia),S2 group(48 cases with osteoporosis).(2)The system of computer-aided diagnosis:(1)region of interest(ROI)extraction: the ROI of 40×40 pixels was manually extracted from the cancellous bone area of the axial CT image of the L3 vertebral(thickness:1mm).Five layers were successively extracted,one ROI was extracted from each layer.In the end,695 ROIs were obtained,including 220 in S0 group,235 in S1 group and 240 in S2 group;(2)Texture analysis: for ROIs,20 texture feature parameters(autocorrelation,contrast,correlation,cluster prominence,cluster shade,dissimilarity,energy,entropy,homogeneity,maximum probability,sum average,variance,sum variance,difference variance,sum entropy,difference entropy,information measures of correlation1,information measures of correlation2,inverse difference normalized,and inverse difference moment normalized)were extracted by gray level co-occurrence matrix(GLCM),each feature parameter includes 4 directions(0°,45°,90°,135°).Thus,a total of 80 feature parameters were obtained for earch ROI;(3)Feature selection: draw the box plot to select effective texture parameters;(4)Classifier recognition: BP neural network,SVM and regression tree classifiers were used to classify the three groups of ROI images;(5)statistical analysis: Matlab was used to draw the ROC curve,and the sensitivity,specificity and area under the ROC curve(AUC)were obtained.Results 1.The results of feature selection: According to the box plots corresponding to the 80 feature parameters,the parameters with poor discrimination were eliminated,and the parameters with better discrimination were retained.Finally,we selected 8 valid parameters,including autocorrelation,contrast,correlation,dissimilarity,energy,entropy,homogeneity,variance.2.Classification results of BP neural network Before the feature selection(80 feature parameters): classifying accuracy rates of S0-S1 was 82.61%,and the AUC was 0.8885.Classifying accuracy rates of S0-S2 was 91.55%,and the AUC was 0.9655.Classifying accuracy rates of S1-S2 was 84.20%,and the AUC was 0.8905.After the feature selection(32 feature parameters): classifying accuracy rates of S0-S1 was 91.07%,and the AUC was 0.9108.Classifying accuracy rates of S0-S2 was 95.82%,and the AUC was 0.9823.Classifying accuracy rates of S1-S2 was 87.62%,and the AUC was 0.9185.3.Classification results of regression tree classifier Before the feature selection(80 feature parameters): classifying accuracy rates of S0-S1 was 81.16%,and the AUC was 0.8555.Classifying accuracy rates of S0-S2 was 91.25%,and the AUC was 0.9280.Classifying accuracy rates of S1-S2 was 86.50%,and the AUC was 0.8920.After the feature selection(32 feature parameters): Classifying accuracy rates of S0-S1 was 88.98%,and the AUC was 0.8897.Classifying accuracy rates of S0-S2 was 94.71%,and the AUC was 0.9659;Classifying accuracy rates of S1-S2 was 87.62%,and the AUC was 0.9235.4.Classification results of SVM classifier Before the feature selection(80 feature parameters): Classifying accuracy rates of S0-S1 was 89.28%,and the AUC was 0.9370.Classifying accuracy rates of S0-S2 was 92.03%,and the AUC was 0.9550.Classifying accuracy rates of S1-S2 was 89.08%,and the AUC was 0.9575.After the feature selection(32 feature parameters): classifying accuracy rates of S0-S1 was 94.01%,and the AUC was 0.9400.Classifying accuracy rates of S0-S2 was 97.21%,and the AUC was 0.9820.Classifying accuracy rates of S1-S2 was 89.15%,and the AUC was 0.9446.Conclusion 1.Normal bone mineral density?osteopenia and osteoporosis could be effectively distinguished by the computer-aided diagnosis based on lumbar spine CT,which provides a method for grading diagnosis of osteoporosis.2.The SVM classifier based on texture feature analysis has a slightly better recognition rate for osteoporosis than BP neural network and regression tree classifier.