Assessment of Vestibular Schwannoma Growth After Radiosurgery: Variability in Clinical Decision Thresholds and the Role of Artificial Intelligence.
Authors
Affiliations (10)
Affiliations (10)
- Yale School of Medicine, New Haven, CT, USA.
- Division of Otolaryngology-Head and Neck Surgery, Department of Surgery, Yale School of Medicine, New Haven, CT, USA.
- Department of Orthopaedics and Rehabilitation, Yale School of Medicine, New Haven, CT, USA.
- Department of Psychiatry, Yale School of Medicine, New Haven, CT, USA.
- School of Medicine and Population Health, University of Sheffield, Sheffield, UK.
- Icahn School of Medicine at Mount Sinai, Elmhurst Hospital Center, Elmhurst, NY, USA.
- Department of Therapeutic Radiology, Yale School of Medicine, New Haven, CT, USA.
- Department of Neurosurgery, Yale School of Medicine, New Haven, CT, USA.
- Department of Radiology and Biomedical Imaging, Yale School of Medicine, New Haven, CT, USA.
- Division of Otolaryngology-Head and Neck Surgery, Department of Surgery, Yale School of Medicine, New Haven, CT, USA. [email protected].
Abstract
To quantify expert interobserver variability in longitudinal vestibular schwannoma growth assessment after radiosurgery and evaluate AI segmentation performance relative to human variability. Retrospective review of patients with sporadic unilateral VS treated with stereotactic radiosurgery (2001-2023). Manual and AI segmentations were compared across the cohort. A growth-enriched subset of 24 tumors underwent additional expert assessment. Agreement was evaluated using intraclass correlation coefficients (ICC), Fleiss and Cohen κ, Bland-Altman analysis, and Dice similarity coefficients. The cohort included 170 MRI examinations from 85 patients (45.9% female; median age, 66 years; median follow-up, 53 months). Human interobserver agreement was excellent for baseline volume (ICC, 0.95), follow-up volume (ICC, 0.98), and percent volume change (ICC, 0.89), but decreased for ≥ 20% growth classification (Fleiss κ, 0.52). Eleven of 24 tumors met the strict-majority growth definition. AI-expert agreement was excellent for baseline and follow-up volumes (ICC, 0.99 and 0.97), but lower for percent change (ICC, 0.83). AI achieved 83.5% binary agreement (Cohen κ, 0.47) and median Dice coefficient of 0.85 (IQR, 0.77-0.90). Bland-Altman analysis demonstrated a mean bias of + 6.64 percentage points (95% limits of agreement, - 53.33 to 66.60). AI-human agreement for absolute tumor volumes was comparable to human interobserver agreement. However, longitudinal growth assessment and binary classification demonstrated greater variability, emphasizing the importance of benchmarking AI performance against human variability when applying clinically relevant growth thresholds.