Outcome prediction of Gamma Knife radiosurgery on vestibular schwannoma using contour-based shape descriptors

Improving care and the transition to personalized medicine is a subject of active research. For many diseases, such as vestibular schwannomas (VS), different treatment strategies are available, each having their own benefits and drawbacks. From a personalized medicine point of view, the ability to predict results for these different treatment strategies would be very useful. In this paper we concentrate on the prediction of the outcome of one of these treatment strategies: Gamma Knife Stereotactic Radiosurgery (GKRS). In some cases, treatment is unsuccessful and the tumor continues to grow. It remains unclear as to why this happens. Some reported influencing factors for a successful GKRS outcome are size and pretreatment growth rate. However, there is conflicting evidence with regard to the predictive value of these factors. Therefore, we have investigated to what extend shape-based binary prediction of GKRS on VS is possible. Twenty-five shape descriptors were computed for training Support Vector Machine (SVM) classifiers and Decision Tree (DT) classifiers. Using these classifiers, 40 tumors of which the treatment outcome is known, were classified. Feature vectors were constructed with 18 descriptors and used for training the classifiers. The best SVM showed a sensitivity, specificity and AUC of 55%, 70% and 0.67, respectively. The best DT classifier resulted in a sensitivity, specificity and AUC of 70%, 60% and 0.71, respectively. These results show that, when using shape-based descriptors, the shape of VS is a weak predictor for GKRS to result into a successful treatment outcome