The Radiation Oncology Feed For radiation oncologists · physicists · RTTs · dosimetrists
technical-relevance brain-metastases cyberknife machine-learning

CyberKnife dose models found size-dependent but non-monotonic brain metastasis control

Tumour-size strata showed non-monotonic BED–control relationships, while machine-learning performance remained unvalidated outside a single retrospective centre.

RadOncToday · ·2 min read

KEY POINTS

  • This single-centre retrospective study included 449 patients with 1,126 lung cancer brain metastases treated using CyberKnife between 2006 and 2025. NSCLC accounted for 85.5% of patients, including 77.7% adenocarcinoma; median follow-up was 15.2 months.
  • Single-fraction SRS delivered 12–24 Gy with a median of 18 Gy, while fractionated SRS delivered 20–40 Gy in two to five fractions, with a median of 30 Gy. Calculated BED₁₀ ranged from 26.4 to 93.3 Gy, and GTV-to-PTV margins were 0–2 mm.
  • Reported median overall survival was 28.9 months, with one- and two-year survival of 72.0% and 53.8%. Median local progression-free survival was 8.8 months, with corresponding one- and two-year rates of 40.1% and 17.8%.
  • Estimated one-year TCD₅₀ values were 57.95 Gy₁₀ for lesions ≤2 cm, 62.52 Gy₁₀ for lesions measuring 2–3 cm, and 59.01 Gy₁₀ for lesions >3 cm (p=0.0161). Two-year estimates were similar at 58.55, 62.48, and 60.11 Gy₁₀, respectively (p=0.0239).
  • Dose–response curves were not consistently biologically plausible: control increased with BED₁₀ for lesions ≤2 cm but decreased at higher BED₁₀ for the 2–3-cm and >3-cm groups. This pattern is highly vulnerable to confounding by indication, as larger or more resistant lesions may have preferentially received higher doses.
  • An XGBoost model predicted a ≥10-point KPS decline within three months with an AUC of 0.874 (95% CI 0.766–0.968), 90.0% accuracy, and an F1 score of 0.933. However, patients dying before the three-month assessment were classified as having no KPS decline, which can materially bias the endpoint.
  • The one-year local-control model reached an AUC of 0.853 (95% CI 0.769–0.926), but some versions incorporated three- and six-month KPS and six-month tumour-control status. The study lacked external validation, driver-mutation data, and systematic radiation-necrosis assessment, while BED₁₀ relied on the standard linear-quadratic model at high fractional doses.

CLINICAL TAKEAWAY

The analysis is useful as an illustration of how tumour size, dose, neurological status, and systemic therapy may interact, but it does not establish individualized CyberKnife prescriptions. The authors’ proposed BED₁₀ ranges should not be implemented: the non-monotonic dose curves, endpoint handling, retrospective confounding, and absent external validation make them unreliable for clinical decision-making.

SOURCE

Advances in Radiation Oncology