AI Segmentation of Vestibular Schwannomas with Radiomic Analysis and Clinical Correlates

Abstract

AbstractVestibular schwannoma (VS) is a benign, slow growing tumor that may affect hearing and balance. It accounts for 7-8% of all primary brain tumors. Gamma knife radiosurgery (GKRS) is a common treatment option for VS. Magnetic resonance imaging (MRI) is employed for diagnosis, surgery planning, and follow-up of VS. Long-term follow-up determines efficacy of VS treatment. Identifying MRI-derived markers to improve management of VS is challenging. This study describes MRI processing pipeline that automatically segments VS and investigates stability and outcome predictive power of radiomic MRI features.We first preprocessed and segmented available pre-GKRS T1-weighted post-contrast MRI images in VS patients, using a Convolutional Neural Network (CNN) developed on DeepMedic framework. Then, we compared CNN and manual segmentations, extracted radiomic features from both manual and CNN segmentations of VS, and, finally, evaluated robustness of extracted features and clinical outcome analyses based thereof.We found that homogeneity, robust maximum intensity and sphericity were the most robust across segmentations. We also found that maximum and minimum intensities were most predictive of tumor growth across all segmentation methods and subject cohorts. We used retrospective post-GK SRS data collected in our institution to build the processing pipeline for unsupervised segmenting of VS. This pipeline is released as a Docker image integrated with XNAT (extensible neuroimaging archive toolkit), an established open research imaging database platform15. Generated segmentations can be viewed and edited in the XNAT-based online OHIF (Open Health Imaging Foundation) viewer16in real time.