Modelling dose and dose‐averaged linear energy transfer to predict high‐grade temporal lobe necrosis following skull‐base proton therapy

Abstract Background In skull‐base proton therapy (PT), severe toxicity outcomes such as temporal lobe necrosis (TLN) may be associated with inadequate management of the actual radiobiological effectiveness (RBE) of proton beams. Combining dose‐averaged linear energy transfer (LET d ) and dose may be crucial in treatment plan optimization and evaluation. Purpose To gather evidence on the combined effect of LET d and dose on voxel‐ and structure‐wise levels in determining a high‐grade TLN (CTCAE v.5 grade ≥G2, G2‐TLN) in skull‐base tumors treated with PT; a dose‐LET d ‐volume histogram (DLVH)‐based model was built to predict G2‐TLN, combined with an analysis of the Dose‐LET d voxel‐based distributions. Methods We retrospectively analyzed data of skull‐base chordomas (61) and chondrosarcomas (18), treated at CNAO National Center for Oncological Hadrontherapy (Pavia, Italy) with PT between September 2011 and July 2020, with a prescription dose of 74 and 70 Gy(RBE), respectively. RBE was set to a constant value of 1.1, while TL dose was constrained to D2cc < 71 Gy(RBE) in the optimization process. Only patients experiencing a G2‐TLN were included in the voxel‐wise analysis and statistically significant differences between dose‐matched LET d distributions within necrotic‐brain areas (necrosis contoured at onset) and healthy‐brain voxels were explored through association tests and mixed‐effects logistic regression models within dose bins. On a structure‐wise level, G2‐TLN association with clinical and DLVH variables was analyzed for each temporal lobe to implement a pre‐processing pipeline. Hence, bootstrap enhanced Elastic‐Net regularized logistic regression, followed by the area under the receiver operating characteristic curve (AUROC), was used to select the final model. Performance and calibration were evaluated with cross‐validation AUROC and Hosmer‐Lemeshow test, respectively. Results With a median follow‐up of 49.7 months, 13 patients experienced G2‐TLN (17 TL) after a median time of 21.5 months. Seven out of 13 G2‐TLN patients reported a statistically significant difference between LET d in necrotic‐ and healthy‐brain voxels after dose‐matching, with 71.4% of them presenting higher LET d values in necrotic‐brain voxels. LET d showed significant associations with the onset of G2‐TLN in dose bins 35–40 Gy(RBE) (Odds Ratio, OR: 7.4, 95%CI: 3.5–26.5, p < 0.001) and 50–55 Gy(RBE) (OR: 2.4, 95%CI: 1.5–4.0, p < 0.001). Regarding structure‐wise modelling, the final logistic regression model was built using two DLVH variables, V (dose,LETd) (V (68,0) and V (19,4.6) ), which showed an independent association with G2‐TLN and proved the highest AUROC. Good performance and calibration were reported by the cross‐validation AUROC (0.89, 95%CI: 0.78–0.95) and Hosmer‐Lemeshow test ( p = 1). Conclusions Voxels receiving higher LET d in the low and medium dose ranges (35–40 and 50–55 Gy(RBE)) were associated with high‐grade necrosis on a voxel‐wise level. The volumes of TL receiving high doses (68 Gy(RBE)) or high LET d (4.6 keV/µm) with doses higher than 19 Gy(RBE) were the major predictors of G2‐TLN. Further evaluation with a larger sample size and an external validation cohort is necessary to assess our findings and validate the presented model.