Classification of computed tomography scanner manufacturer using support vector machine

Abstract

Computed tomography (CT) is useful to investigate the presence and severity of injury during acute stage of traumatic brain injury (TBI) due to its availability and short image acquisition time. Recently, quantitative CT analysis have shown promising results in objective and accurate assessment of lesion and the prediction of outcome. To conduct further multicenter, longitudinal follow-up studies using quantitative analysis, the effect of CT scanner manufacturer should be investigated. In this study, CT images were acquired from 326 subjects without any apparent intracranial abnormalities. The images were scanned by three different scanner manufacturers. The quantitative analysis was performed and plotted as density distribution. The acquired density distributions were served as input features of support vector machine (SVM) using Gaussian kernel function, which is designed for classifying the CT images based on the scanner manufacturers. The optimal hyperparameters were explored via grid search test and the model increased the robustness by 5-fold cross validation. The best predictive performance was obtained when C = 100 and γ = 0.1 (accuracy = 91.1 %). The results showed significant difference in density distribution according to the scanner manufacturers, and thus suggest that the manufacturer should be standardized to conduct the quantitative analysis on the brain CT images.