ESTRO 2024 - Abstract Book

S4615

Physics - Optimisation, algorithms and applications for ion beam treatment planning

ESTR0 2024

dataset. The model was externally validated on 28 PBS patients from the West German Proton Therapy Centre Essen. Evaluations of the prediction quality were performed on two levels.

1. Quantitative analyses of the LETd distributions were performed in CTV, brainstem, chiasm and optic nerves (left / right). To assess differences between MC- and DL-based LETd distributions, the Wilcoxon signed-rank test was used. In addition, the voxel-wise Root Mean Squared Error (RMSE) between MC- and DL-based LETd distributions was calculated within the regions. 2. The impact of the LETd prediction error on normal tissue complication probability (NTCP) models was assessed for several relevant side-effects. The RBE-weighted dose [6] was calculated using the MC- and the predicted DL-based LETd distribution. Based on the RBE-weighted dose, treatment-related side-effects (acute ocular toxicity grade ≥2 and memory impairment grade ≥1 after 24 months) were estimated using published NTCP models [7,8]. The difference in NTCP between MC and DL was quantified by the RMSE.

Results:

The SegResNet performed best in internal validation and showed good external validation performance with an RMSE for the mean LETd in the CTV, brainstem, chiasm and optic nerves (L/R) of 0.25, 0.22, 0.33, 0.48, 0.68 and 0.35 keV/μm, respectively, and RMSE ≤ 0.9 keV/μm for the 98 percentile in all regions. No statistically significant differences were found for any of the healthy tissues (Figure 1). Side-effect predictions showed minor deviations when comparing LETd from DL and MC models (RMSE < 0.02 for all considered side-effects) (Figure 2).

Figure 1: Mean (first row) and 98 percentile (second row) of the voxel-wise LETd distribution for patients of the external validation data in selected regions of interest based on Monte Carlo (MC) simulation (blue) and deep learning (DL)-based calculation (orange). P-values marked with * show statistically significant differences. Green dashed lines indicate distribution means.