| This project approached the old problem of estimating the severity of scoliotic deformity from torso surface with Support Vector Machine (SVM), a type of novel learning methodology. The main differences between our method and Jaremko's were that firstly, we did not extract any feature from torso surface; secondly, we employed SVM instead of ANN in order to have better generalization performance. The scanned data points were represented by the control points of the surface that fitted to these scanned raw points. A linear dimension reduction technique, principal component analysis, was also involved.; We tested our method on two types of datasets: Brace dataset and Calgary dataset (containing 41 and 115 data, respectively), on which we committed the classification experiments. The patients were divided into several classes according to their Cobb angle. On Brace dataset, two classes were defined by using the mean Cobb angle of all patients in the dataset as the threshold. On Calgary dataset, three classes (Cobb angle <30°, 30--50°, and >50°, respectively) were defined which corresponded to patients having mild, moderate, and severe spinal deformity, respectively. The best results obtained on Brace dataset were 0% training error and 29.27% test error with ERBF kernel. The best results obtained on Calgary dataset were 9.71% training error and 32.97% test error with RBF kernel. Comparing to the result of GA-ANN method, this result is relatively poor. We think the main cause is from the nature of control points that are not a steady and invariant representation of surface deformities, and it does not capture the deformity information as accurately as features. (Abstract shortened by UMI.)... |
