2164 - 18F-Faza PET Radiomics for Prediction of Temporal Lobe Radiation Necrosis after Proton Therapy In Skull Base Chordoma

03:00pm - 04:00pm ET

Poster Hall - Exhibit Hall A

Screen: 10

POSTER

Presenter(s)

Hamid Mammar, MD, PhD, MD - Institut Curie, Paris, Ile de france

B. Nicaise¹, A. Bonfrate¹, P. Loap¹, E. Le Reun¹, P. Isabelle¹, F. Goudjil¹, L. Feuvret², J. Jacob³, S. Helfre¹, V. Calugaru¹, R. Dendale¹, L. De Marzi¹, and H. Mammar¹; ¹Institut Curie, Paris, France, ²Hospices de Lyon, Lyon, France, ³University Hospitals Pitié-Salpêtrière, Paris, France

Purpose/Objective(s): High-dose proton therapy (HDPT)has improved local control in skull base chordomas (SBC) but has also increased the incidence of temporal lobe radionecrosis (TLRN). 18F-FAZA is a radiopharmaceutical used as a biomarker of tissue hypoxia. This study investigates the association between pre-treatment 18F-FAZA PET radiomic features and the occurrence of TLRN

Materials/Methods: Thirty Two patients with SBC enrolled in the Protonchorde-01 trial and treated with HDPT between 2017 and 2020 were analyzed. Following gross tumor resection, a dose of 70.2 Gy(RBE) was delivered to the surgical bed, with a boost up to 78.2 Gy(RBE) to hypoxic areas. Six patients (18.7 %) developed TLRN after treatment.

18F-FAZA PET/CT imaging was performed postoperatively and prior to PT as part of the study protocol. The injected activity was 5 MBq/kg (±20%), and image acquisition occurred 180 minutes post-injection. Temporal lobes were automatically segmented on CT images co-registered with PET using TotalSegmentator v2.12.0. First-order radiomic features based on PET intensities within the temporal lobes were extracted using PyRadiomics v3.1.0. Radiomic features were compared between patients with and without TLRN using the Mann–Whitney U test (SciPy v1.15.2). Logistic regression models with L2 regularization were trained using statistically significant features and evaluated with 5-fold stratified cross-validation (scikit-learn v1.7.2). Performance metrics and 95% confidence intervals were estimated using 2,000 bootstrap resamples

Results: The 10th percentile and median intensity are significantly associated with TLRN (p=0.036 and 0.044, respectively). Logistic models using these features achieved ROC AUCs between 0.76 and 0.80, with balanced accuracies ranging from 0.90 to 0.93. Odds ratios indicate that lower intensity values are associated with a higher occurrence of TLRN.

Conclusion: The 10th percentile and median PET intensities were significantly associated with the development of TLRN. These findings suggest that higher temporal lobe oxygenation may be associated with the occurrence of TLRN. Further investigations in a larger cohort and incorporating pre-treatment multimodal MRI are ongoing