Proton brain injury model supports increased periventricular radiosensitivity and LET-dependent risk

EXCERPT In 112 proton-treated low-grade glioma patients, 21% developed contrast-enhancing brain lesions, with modeling supporting increased periventricular and LET-dependent injury risk.

KEY POINTS

  • This retrospective study from Heidelberg Ion-Beam Therapy Center independently evaluated an existing voxel-based prediction model for radiation-induced contrast-enhancing brain lesions (CEBLs) after proton therapy for low-grade glioma. The validation cohort included 112 patients treated between 2015 and 2019, compared with the original model-development cohort of 110 patients. Median age was 37 years, median prescription dose was 54 Gy(RBE), and 86% received exactly 54 Gy(RBE), generally using 1.8–2.0 Gy(RBE) per fraction.
  • The incidence of new contrast-enhancing brain lesions was 21% in both cohorts. In the validation population, 23/112 patients developed at least one CEBL, with 26 lesions identified compared with 67 lesions in the original cohort. Median lesion volume at first detection was 110 mm³ versus 235 mm³, respectively. Median follow-up was 23 months, compared with 39 months in the development cohort. Lesions were defined as new contrast enhancement in normal brain tissue outside the gross tumor volume.
  • The original model incorporated physical dose, dose-averaged LET, and proximity to the ventricular system. Periventricular tissue was defined as a 4-mm expansion around the cerebral ventricles. Patient-specific treatment plans were recalculated with Monte Carlo methods to determine physical dose and dose-averaged LET distributions. Voxel-level probability-of-lesion-origin estimates were combined into a patient-level normal tissue complication probability under an assumed serial injury model.
  • Overall patient-level calibration was favorable, but regional predictions were inaccurate. The original model predicted an average CEBL incidence of 20%, closely matching the observed 21%. However, predicted periventricular lesion risk was 6% versus 13% observed, while predicted risk in the remaining brain was 15% versus 11% observed. Of the 26 observed lesions, 14 occurred within the periventricular region and 12 elsewhere, indicating that good overall calibration concealed substantial spatial miscalibration.
  • Updating the model strengthened both the periventricular and LET-associated risk terms. The periventricular coefficient increased from 1.19 to 1.99, corresponding to an estimated voxel-level odds ratio of approximately 7.3 for lesion origin in periventricular tissue. The physical-dose coefficient decreased from 0.19 to 0.13 Gy⁻¹, while the dose–LET interaction coefficient increased from 0.018 to 0.024 (Gy·keV/μm)⁻¹. Following recalibration, predicted regional risks were 10% for periventricular tissue and 11% elsewhere, compared with observed rates of 13% and 11%.
  • The updated voxel-level model achieved high discrimination, but its accuracy requires careful interpretation. Reported voxel-level AUC was 0.95, with a Brier score of approximately 7.7 × 10⁻⁵. However, CEBL-positive voxels represented an extremely small fraction of all analyzed brain voxels, and voxel-based resampling may underestimate uncertainty because neighboring voxels from the same patient are correlated. Patient-level resampling produced substantially wider coefficient confidence intervals, although the mean estimates remained broadly stable.
  • The study provides clinical imaging evidence supporting variable proton biological effectiveness, not a directly measured RBE value. Validation was conducted using patients treated more recently at the same institution rather than a completely independent external center. The cohorts differed in target volumes, chemotherapy exposure, treatment field configurations, and follow-up duration. CEBLs are also an imaging endpoint rather than synonymous with symptomatic radionecrosis. The findings support investigating LET-aware planning and periventricular sparing, but do not establish validated dose constraints or prove that model-guided planning prevents clinically significant injury.

CLINICAL TAKEAWAY

This independent temporal validation supports the association of proton dose, LET, and ventricular proximity with radiation-induced contrast-enhancing lesions after low-grade glioma treatment. The findings strengthen the rationale for treating periventricular tissue as a potentially sensitive structure and evaluating LET distribution during proton planning. However, prospective multicenter validation and evidence that risk-adapted planning reduces clinically meaningful toxicity are still needed.

SOURCE

Radiotherapy and Oncology

Browse more research Suggest a correction