1123 - WEAVE: A Clinical Support Platform for Automated Volumetric Analysis and Longitudinal Management of Vestibular Schwannoma
Presenter(s)
G. A. Szalkowski1, R. Prashad1, J. S. Fernandes2, F. Lam3, A. Kattaa3, C. F. Chuang1, L. Wang4, L. Liu2, M. Chen5, Q. Wang5, S. G. Soltys1, E. L. Pollom1, E. Rahimy1, D. Park3, Y. Hori3, H. Jiang5, W. Lu5, X. Gu6, and I. C. Gibbs1; 1Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA, 2Department of Radiation Oncology, Stanford University, Stanford, CA, 3Department of Neurosurgery, Stanford University School of Medicine, Stanford, CA, 4Department of Radiation Oncology, Stanford University, Palo Alto, CA, 5Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Dallas, TX, 6Stanford University Department of Radiation Oncology, Palo Alto, CA
Purpose/Objective(s): Management of vestibular schwannoma (VS) relies on precise longitudinal volumetric assessment to guide "wait-and-scan" surveillance, microsurgery, or radiosurgery planning. Manual segmentation is labor-intensive and subject to interobserver variability. We developed the WEb-Accessible comprehensiVE (WEAVE) platform, an AI-driven clinical tool designed to automate VS segmentation and provide standardized longitudinal tracking of tumor volumes to streamline clinical workflows and enhance treatment monitoring.
Materials/Methods: The WEAVE platform integrates a high-performance web interface with a robust backend auto-segmentation engine and database. The front-end web client is built on Rust and WebAssembly featuring fast image rendering on web browsers. The back-end server is established on a SQLite database and an AI-autosegmentation engine. Recognizing varied clinical imaging protocols, we utilized a nnU-Net architecture to develop three distinct models optimized for different modality combinations: (1) CT/T1c/T2, (2) T1c/T2, and (3) T1c-only. Model performance was validated against expert manual delineations (ground truth) on 20 test cases. Clinical utility was assessed using metrics including Dice Similarity Coefficient (DSC), Absolute and Relative Volume Differences (AVD/RVD), and 95th percentile Hausdorff Distance (HD95) to evaluate boundary accuracy.
Results: All three models demonstrated high clinical fidelity with no significant performance variance. Mean DSC scores reached 0.89–0.9; this quality is acceptable for clinical use. The average AVD was minimal (0.11–0.13cc), with mean HD95 values below 1.0mm (0.74–0.88mm), ensuring sub-millimeter boundary precision. Automated inference was completed in approximately 60 seconds per case. The WEAVE platform enabled seamless web-based visualization, providing physicians with longitudinal tumor growth curves and side-by-side comparisons of historical scans to facilitate rapid assessment of treatment response or disease progression.
Conclusion: The WEAVE platform offers a clinically adequate solution for automated VS management, bridging the gap between sophisticated AI segmentation and bedside utility. By achieving accuracy levels on par with manual contouring while significantly reducing processing time, the platform enhances consistency in SRS/SRT planning and surveillance. This framework provides a scalable model for automating the management of other complex intracranial pathologies, potentially improving the precision of multi-disciplinary neuro-oncology care.