Classification of osteoporosis by extracting the microarchitectural properties of trabecular bone from DXA scans based on thresholding technique

The purpose of this study was to find optimal thresholding conditions for diagnosing osteoporosis using a novel thresholding technique. Seven trabecular features-composed of four structural and two skeletonized features as well as the fractal dimension (FD)-were extracted from 2D DXA scans. A binarized image was used to identify the structural features and the FD, and the skeletonized feature sets were obtained from a skeletonized image. The proposed thresholding technique utilizes the percentages for the trabecular bone area as the threshold values after dividing them into ranges from 0% to 95%. Principal component analysis was used to identify the optimum subsets from the original trabecular features. A two-layered feed forward neural network used a Bayesian regularization training algorithm to carry out the classification. In the results of the experiment, the proposed method exhibited superior performance in discriminating osteoporotic patients from normal individuals. The overall accuracy was significantly higher, in the ranges of 10-40%, and a threshold condition of 20% presented the highest classification performance when considering all of the experimental conditions. The accuracy of the classification gradually decreased as the threshold increased beyond 20%. These results suggest that the proposed method has clinical applicability for diagnosing osteoporosis.