Computer-aided prediction of tumor response one year after Gamma Knife radiosurgery on vestibular schwannoma using deep learning

Introduction Approximately one third of all vestibular schwannomas (VS) show a volumetric increase one year after Gamma Knife radiosurgery (GKRS) treatment, which can indicate transient tumor enlargement (TTE) and may cause symptoms due to increased mass effect. It is therefore clinically relevant to be able to predict tumor response relatively early after GKRS. Computer-aided prediction models using classical machine learning have already shown promising results in predicting GKRS tumor response in VS patients. The objective of this study is to investigate the feasibility of the modern machine learning method of deep learning to predict early tumor response after GKRS. Methods A total of 1118 contrast-enhanced T1 MRI scans, obtained during GKRS treatment of VS patients, were used. These images were cropped to include only the tumor and its direct surroundings using a bounding box. This data was subsequently used to train a three-dimensional autoencoder model to extract 256 tumor image features automatically. The learned features were subsequently employed in a neural network classifier to predict tumor response based on the treatment scan. For this classification task, tumor enlargement and shrinkage were defined as at least a 15% volumetric increase and decrease with respect to the tumor volume at the time of GKRS. Furthermore, patients without a one-year follow-up scan were excluded, as well as patients with tumors smaller than 0.25cc due to the lack of sufficient image features. Results The exclusion criteria resulted in the inclusion of 395 patients with a mean one-year follow-up time of 12.14±0.70 months. Using our volumetric criteria, enlargement was observed in 122 (30.9%) tumors and shrinkage in 273 (69.1%). The prediction model achieved a sensitivity and specificity of 72.3% and 82.5%, respectively. Conclusion The specificity shows that the model is more tuned to predicting tumor shrinkage. Although tumor enlargement prediction performance (i.e., sensitivity) is lower, the overall performance of the model indicates that deep learning can be of predictive value for tumor response in VS patients. These methods should therefore be further investigated to improve upon the presented results