Prediction of Vestibular Schwannoma Surgical Outcome Using Deep Neural Network

Abstract

To compare shallow machine learning models and deep neural network (DNN) model in prediction of vestibular schwannoma (VS) surgical outcome. 188 patients with VS were included, all underwent suboccipital retrosigmoid sinus approach, and preoperative MRI recorded a series of patient characteristics. Degree of tumor resection was collected during surgery and facial nerve function was evaluated on the 8th day after surgery. Potential predictors of VS surgical outcome were obtained by univariate analysis, including tumor diameter, tumor volume, tumor surface area, brain tissue edema, tumor property and tumor shape respectively. This study proposes a DNN framework to predict the prognosis of VS surgical outcomes based on potential predictors, and compares it with a series of classic machine learning algorithms including logistic regression. The results showed that three predictors of tumor diameter, tumor volume, and tumor surface area were the most important prognostic factors for VS surgical outcomes, followed by tumor shape, while brain tissue edema and tumor property were the least influential. Different from shallow machine learning models, such as logistic regression with average performance (area under the curve (AUC): 0.8263; accuracy: 81.38%), the proposed DNN shows better performance, where AUC and accuracy were 0.8723 and 85.64% respectively. Based on potential risk factors, DNN can be exploited to achieve preoperative automatic assessment of VS surgical outcomes, and its performance is significantly better than other methods. It is therefore highly warranted to continue to investigate their utility as complementary clinical tools in predicting surgical outcomes preoperatively.