Preliminary Study On Classification Of Osteoporosis Diagnosis From X-ray Plain Film Using Computer Aided Diagnosis System

Purpose: The present work is an effort to explore whether back-propagation(BP)neural network is useful for improving osteoporosis diagnostic efficacy by analyzing the trabecular bone texture in pelvic radiographic images by gray level co-occurrence matrix.Methods: Of 143 patients who visited the second affiliated hospital of Dalian medical university for pelvic radiographic image between January and December in 2014,44 patients with DXA measurement were recruited for this study.These patients were divided into 3 groups according to their BMD.15 patients presented S0(normal BMD),15 S1(osteopenia),and 14 S2(osteoporosis).Firstly,2-4 regions of interest(ROI)of 20�20 pixels were selected by a trained operator at the bilateral femoral neck Ward's triangle regions.328 ROIs were totally selected,116 ROIs for S0,108 ROIs for S1,and104 ROIs for S2.Secondly,a computer aided diagnosis system was built with Matlab based on texture analysis.Then gray level co-occurrence matrix was performed to analysis 20 texture parameters in 4 dimensions(0�,45�,90� and 135�),including 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.Optimization of texture parameters was performed according to 80 box plots of texture features.Thirdly,the back-propagation(BP)neural net work was used to classify S0-S1,S0-S2 and S1-S2.Lastly,sensitivity and specificity were calculated by Matlab with ROC curve performed.Results:1.Classification results of BP classifier based on 80 texture features Classifying accuracy rates of S0-S1 was 76.79%,and sensitivity and specificity were79.16% and75.00% respectively.Classifying accuracy rates of S0-S2 was76.36%,the sensitivity and specificity were 74.07% and 78.57% respectively.Classifying accuracy rates of S1-S2 was 60.85%,the sensitivity and specificity were 60.19% and 61.47%respectively.2.According to 80 box plots of texture features,invalid texture features were eliminated for optimization.The remaining 11 texture features were used for texture extraction and analysis,including autocorrelation,contrast,dissimilarity,homogeneity,sum average,variance,sum variance,difference variance,sum entropy,difference entropy,and inverse difference normalized.3.Classification results of BP classifier based on 44 texture features Classifying accuracy rates of S0-S1 was 78.57%,the sensitivity and specificity were78.84% and 73.31% respectively.Classifying accuracy rates of S0-S2 was 78.64%,the sensitivity and specificity were 79.44% and 80.53% respectively.Classifying accuracy rates of S1-S2 was62.74%,the sensitivity and specificity were 61.47% and 64.08%respectively.Conclusion:1.S0-S1 and S0-S2 could be effectively distinguished by the computer aided diagnosis system based on texture analysis from pelvic X-ray plain film.2.The computer aided diagnosis system based on texture analysis provided a new method for classification of osteoporosis diagnosis.