The RBF kernel was used to process 60 training sets and classify 40 test sets. For the training sets, the data was classified into an osteoporotic and a normal group and labelled as either '0' or '1' that corresponded to positive and negative examples for SVM training. The RBF kernel-SVM predictions for the diagnostic classification of women had 90.9% sensitivity and 83.8% specificity on the basis of the lumbar spine BMD and 90.0% sensitivity and 69.6% specificity on the basis of the femoral neck BMD (Table ). The level of performance of the classifications was high except for the low specificity obtained for femoral neck BMD data. The accuracy of the proposed CAD system with SVM for diagnosing women with low BMD at the lumbar spine was 88%, PPV was 71.4% and NPV was 96.7%; on the basis of the femoral neck BMD, the accuracy was 75%, PPV was 46.6% and NPV was 96.0% as presented in Table . The overall accuracy of the trimmed mean CAD method [13
] for the classifications on the basis of the lumbar spine BMD, was 78%, PPV was 47.4% and NPV was 96.8%, whereas on the basis of the femoral neck BMD, the accuracy was 72%, PPV was 42.9% and NPV was 93.1%. In addition, the sensitivity and specificity for the combined data of both the lumbar spine and femoral neck BMD using SVM method processed with 130 training and 70 testing sets were 90.6% and 80.9% respectively (Table ). The average time to complete the classification measurement on a single radiograph was only 9.0 s. The accuracy of the classifications for 10 randomly selected subjects that were measured twice with a one-month interval was 90%.
Performance of the proposed SVM method for identifying women with low lumbar spine BMD and femoral neck BMD at a 95% confidence interval (CI)
Performance of the proposed SVM method for identifying women with combined skeletal bone mineral densities (BMD) at a 95% confidence interval (CI)
We used our proposed SVM method with a CAD system and dental panoramic radiographs to diagnose women with low BMD easily and quickly. The use of the SVM kernel in this study provided a high degree of consistency and reproducibility in the results. One of the key advantages of this CAD system over a manual assessment is the objectivity of the automated evaluation. The proposed CAD system directly assesses the bones on radiographs by measuring the MCW continuously between the mental foramen and mandibular angle, which can reduce measurement errors that occur with the conventional assessment [11
]. The trimmed mean method (accuracy = 79%) of the recently proposed CAD method [13
] can be replaced by the SVM method (accuracy = 88%) to obtain a better result. The proposed SVM diagnostic model performs differential diagnosis very well. Because classification by the trimmed mean method requires prior threshold setting, the SVM approach for classifying women as having either a low or normal BMD may be a better choice.
In the previously developed CAD systems [11
], the diagnostic accuracy of the radiologist was adversely affected by manual interactions. Arifin et al. [11
] reported that measurement of the MCW at one point (below the mental foramen) by CAD had a detection sensitivity of 88% and specificity of 58.7%. The conventional method of continuous measurements using the trimmed mean technique [13
] was reported to have a sensitivity of 90% and specificity of 75% on the basis of the lumbar spine BMD and a sensitivity of 81.8% and specificity of 69.2% on the basis of the femoral neck BMD. However, for our newly proposed SVM method, the sensitivity and specificity were 90.9% and 83.8%, respectively on the basis of the lumbar spine BMD and 90.0% and 69.6%, respectively on the basis of the femoral neck BMD.
These findings indicate that the classification performance of the diagnostic system using the SVM method achieved higher accuracy for detecting women with low BMD compared with the conventional approach. The difference between these results is reasonable because unlike other techniques, the CAD system with SVM has an advantage of converging the problem to the global optimum and not to a local optimum.
Our findings are supported by those of previous studies. Lim et al. [28
] evaluated bone fractures from X-ray images with different classifiers and reported a high classification accuracy of 98.2% from a method using a combination of SVM classifiers. Caligiuri et al. [29
] showed that their method was promising for discriminating between healthy and fractured bones with high Az
values. It was also reported that the sensitivity and specificity for an RBF kernel-SVM that was used for the reorganization of nuclear receptors was almost equal to our classification system [30
]. Comparisons of our experimental results with those of previous studies demonstrated the feasibility and excellent performance of our proposed system in diagnosing high-risk groups with low BMD or osteoporosis. Several screening tools based on simple questionnaires have been developed to identify postmenopausal women with low skeletal BMD or osteoporosis, and validation of these tools has also been performed in many countries [31
]. The sensitivity and specificity of such decision rules in identifying postmenopausal women with osteoporosis ranged from 90% to 92% and 37%-45%, respectively. This proves that the diagnostic efficacy of our SVM method is better than that of the several questionnaire-based screening tools that were used in the previous studies, although the backgrounds of the subjects were different. The limitations in our study were that the number of subjects was relatively small, and the subjects were relatively healthy postmenopausal women because we used rigid exclusion criteria.