2026 Poster Presentations
P504: BICENTRIC STUDY OF AI-BASED DIFFERENTIATION OF MALIGNANCY IN SKULL BASE SCHWANNOMAS AND PROGNOSTIC PREDICTION
Daniela Stastna, MD, MPhil1; Danielle Dang, MD1; Robert Macfarlane, MD2; Richard Mannion, MD, PhD2; Patrick Axon, MD3; James Tysome, MD, PhD3; Neil Donnelly, MD3; Carolina Giannini, MD, PhD4; Robert Spinner1; Jamie Van Gompel, MD, PhD1; Michael Link, MD, PhD1; 1Neurosurgery, Mayo Clinic, Rochester, MN; 2Neurosurgery, University of Cambridge Hospitals, Cambridge, UK; 3ENT, University of Cambridge Hospitals, Cambridge, UK; 4Pathology, Mayo Clinic, Rochester, MN
Introduction, Aim of the study: Malignant peripheral nerve sheath tumors (MPNST) are locally aggressive tumors, rarely affecting cranial nerves [CN VIII>V>other]. Malignant transformation is reported in less than 1% of CN schwannomas, and associated predominantly with radiation induction or neurofibromatosis type 1 (NF1).
Prognosis in MPNSTs is poor, with rapidly progressive tumor volume and morbidity; 5-year survival of 30–60%.Therefore, distinguishing benign schwannomas, often managed with surveillance, from MPNSTs is critical.
This bicentric retrospective study aims to:(1) predict malignant characteristics of MPNST using machine learning(ML) on baseline imaging(MRI), (2) identify predictors of malignancy, (3) estimate long-term outcomes. To the best of our knowledge, this represents the first study of its kind.
Methods: A bicentric retrospective study was conducted on 10 patients who underwent resection of cranial nerve MPNST (CN-MPNST). Eleventh patient with cerebral MPSNT metastasis was excluded. Survival analysis (Kaplan–Meier with log-rank, Cox_proportional_hazards_regression) were applied to assess progression-free survival (PFS) and overall survival (OS) of 10 CN-MPNST compared to 209 resected benign schwannomas (209/610).
Radiomic features were extracted from pre-operative contrast-enhanced CET1-weighted MRIs, and used to train a XGBoost and four-layer sequential neural network (SNN,Python/Keras) models against a comparative dataset of 600 benign vestibular (CN VIII) and 10 trigeminal (CN V) schwannomas, and 26 peripheral MPNST. Model performance was evaluated using AUC-ROC, training/validation accuracy, and loss.
Results: Our bicentric study included 10 patients with CN-MPNST, operated between 1985- 2025. The median age was 57 years (range36–66), with 60% males.
Regarding the etiology of CN-MPNST, 70% were radiation-induced (occurring 6.5–26years after radiotherapy), while 30% represented de novo transformation (3/10). NF1 was confirmed in one radiation-induced case. The most frequently affected cranial nerves were the CN VIII( 5/10) and the CN V (3/10).
Gross-total resection (GTR) was achieved in 60% of patients (6/10), with adjuvant radiotherapy administered in 90%. The GTR rate was lower in CN-MPNST compared with benign schwannomas (60% vs. 71%). Histopathology showed elevated Ki-67_index (median 50%,range 45–90) and loss of H3K27me3 in radiation-induced cases(5/7).
Median PFS among CN-MPNST was 0.6 years (95% CI,0.1–3.6). Overall mortality was 60%, with a median OS of 1.3 years from surgery (95% CI,0.2–3.6). Incomplete resection(STR/NTR) was a significant predictor of worse PFS and OS.(Figure1). GTR significantly prolonged PFS (2.6 years,95% CI,1.5–3.6) compared with STR/NTR (0.3 years; p = 0.0033;HR=25.3).Radiation-induced etiology, elevated Ki-67, and involvement of CN VII/X had associations with unfavorable PFS/OS, but nonsignificant.
Our SNN achieved accurate distinction of MPNST on preoperative CET1-wMRI from 610 benign schwannomas (Weighted_accuracy: 0.933, Mean_Squared_Error(MSE): 0.0762, F1-score:0.97).(Figure 2) XGBoost model along with features’ SHAP values achieved accuracy of 0.97, AUC:0.99. (Figure 3). The most significant features from both models identified irregular shape and complex signal intensity (heterogeneity undetectable to the human eye) as the most distinctive features for MPNST.
Conclusions: Understanding the histopathological, genetic, and radiological features of CN-MPNST is essential for accurate diagnosis, personalized treatment planning, and prognostication. Early detection of malignant transformation using routine MRI might enable timely interventions, and potentially improve patient outcomes.
.jpg)
