Machine Learning in Vestibular Schwannomas: Automated Detection, Segmentation and Prediction of Future Growth of Postoperative Residuals

Introduction: The extent of surgery in sporadic vestibular schwannomas(VS) is one the most significant predictors of recurrence.However,distinguishing tumor residuals from postoperative changes is often challenging.Our goal was to develop a reliable algorithm capable of differentiating tumor residuals from postoperative changes,and predicting future growth.

Methods: This retrospective study is based on adult patients who underwent primary incomplete surgery for Koos III-IV sporadic VS at our center, between I/2008-XII/2018. Pre-/postoperative CET1-weighted MRI were analysed using our algorithm(Python, Keras). Preoperative semi-automated segmentations served as ground truth for transfer learning. Due to a small number of cases, the combination of complex features with sequential neural network(SNN) classification were used.

The 1st part :Our SNN/3-layered_perceptron determined the most significant features, which were used as thresholds for Chan-Vese_algorithm in automatic detection,3D-segmentation of residual. DICE score assessed agreement between “predicted residue”/ “ground truth” residuals. Moreover,SNN model performed the distinction between real residuals and “postoperative” changes.

The 2nd part: Another models (XGBoost,SNN) were applied to predict the future behavior(5-year_PFS) of the residual.

Results: Among 119 patients with incomplete resection,67 patients had visible residuals on the baseline MRI.We augmented dataset using geometric transformations of 13 random patients,creating cohort of 80 patients.

Part 1: Automated Detection_Segmentation of Residuals
The initial "seeding 2x2 voxels" were defined as the intersections of the Chan-Vese,feature-based map within predefined FOV "bilateral_IAC". This feature-based map was thresholded using the mean±SD of the best 3 features from SNN model. The overall DICE score for the 80 residuals was 0.92(strong agreement).Moreover,80 residuals were tested against 40 “postoperative” changes. The SNN model achieved validation accuracy of 0.933 for distinguishing residuals from postoperative changes.

The 2nd part:
XGBoost model after tuning achieved MSE=0.03, AUC=0.97,with accuracy 0.95(95%CI 0.86–1.05),precision-recall=0.83. The SNN/MPL model achieved validation accuracy of 0.89.

Conclusion : Despite a small dataset, the algorithm combining transfer learning, complex feature extraction with SNN achieved reliable performance in detecting VS residuals, as predicting regrowth.