ESTRO 2024 - Abstract Book

S2032

Clinical - Paediatric

ESTRO 2024

features. This score was dichotomized using a time-dependent ROC curve to classify patients into low-risk and high-risk groups. Difference in the Kaplan-Meier (KM) curves between these groups was assessed with the log rank test. Univariate analysis was performed using the log-rank test for qualitative variables or the Cox proportional hazards model for continuous variables. Multivariate analyses were performed using Cox proportional hazards model. One model was built with relevant clinical characteristics. A backward stepwise selection was used to determine the best model adjusted on visually-derived characteristics. An integrative model, which combined the radiomic signature (Rad-score), clinical and visually-derived characteristics was also built. The concordance index (C-index) of each model was used to quantify their performance.

Results:

The median follow-up was 67.7 months (95% confidence interval [CI] [58.4; 73.9]). Median age at start of RT was 4 years ((1.0; 22.0). Regarding treatment, 86.5% of patients underwent gross total resection, 65.4% received photon therapy (XRT) vs. 33.2% proton beam therapy (PBT) and 1.4% for a mix of PBT and XRT, 59.7% received a dose ≥59.4 Gy. Of the 87 relapses, 63.2% were local.

Radiomics analysis identified 10 features with value in evaluating RFS (2 first-order and 8 second-order).

Univariate analysis showed that among clinical and visually-derived characteristics, low RT dose (p=0.008), young age at start of RT (p=0.006), age<4y (p=0.008), second look or more resections before RT (p=0.024) and residual tumor (p=0.044) were significantly associated with worse RFS.

We found association between radiomic signature (Rad-score built using the 10 selected features) and RFS (hazard ratio [HR] = 1.23 [1.12; 1.36], p<0.001) with a C-index = 0.63.

Based on the time-dependent ROC curve analysis, the EPN patients were stratified into high- and low-risk groups (70.6% vs. 29.4%) and the difference between KM curves of these two groups was significant (p=0.005).

In multivariate analysis, five clinical characteristics (delay time between diagnosis and RT, age at start of RT, RT dose, radiation technique and extent of resection) and one visually-derived characteristic (necrosis) were selected for clinical and visual models building respectively. Only young age at start of RT (HR=0.94 [0.90; 0.99], p=0.027) was a prognostic factor of worse RFS, adjusted for other variables, with a C-index of 0.62.

Finally, the integrative model evaluated RFS (C-index 0.68) better than either the Rad-score (C-index 0.63) or the clinical (C-index 0.62) and visual models (C-index 0.54).

Conclusion:

The preliminary results of this study demonstrated that combining the radiomic signature, clinical, and visually derived characteristics showed the best potential in predicting RFS for ependymoma patients. The MRI radiomic signature may confer incremental value over clinical and visual features. Future work will consist in integrating molecular classification, evaluating the potential benefit of harmonizing MR images with deep learning-based image synthesis before features extraction, and performing a proper validation of the developed models in an external testing cohort.

Keywords: Ependymoma, relapse, radiomics