Is Real-Time Inverse Planning Optimizing Dose to the Normal Brain? A Prospective Comparative Trial in a Series of Brain Metastases Treated by Stereotactic Radiosurgery

Abstract

Background: Radiosurgery has demonstrated good safety and efficacy in the treatment of multiple brain metastases (BMs). However, multi-target dose planning can be challenging and time-consuming. A recently developed real-time inverse treatment planning (IP) by convex optimization has been demonstrated to produce high-quality treatment plans with good conformity and selectivity in single-target plans. We intended to test the capacity of this IP to rapidly generate efficient plans while optimizing the preservation of normal tissue in multiple BM. Methods: Seventy-nine patients (mean age 62.4, age range 22–85) with a total of 272 BMs were treated by Gamma Knife Radiosurgery. All subjects were treated using a forward planning (FP) technique by an expert neurosurgeon. The new Intuitive Plan was applied and able to automatically generate an alternative plan for each patient. All planning variables were collected from the IP to be compared with the corresponding measurements obtained from the FP. A paired sample t test was applied to compare the 2 plans for the following variables: brain volumes receiving 10 Gy (V10) (primary endpoint), and 12 Gy (V12), planning indices (selectivity, coverage, gradient, and Paddick Conformity Index [PCI]), beam-on time (BOT), and integral doses. Additionally, the noninferiority margin for each item was calculated, and the 2 plans were compared for noninferiority using a paired t test. Results: The mean age of patients was 62.4 years old (age range 22–85), with a sex ratio of 1.02. The average number of lesions per patient was 3.4 (range 1–12). The mean prescription dose was 21.46 Gy (range 14–24 Gy). Noninferiority of the IP was concluded for V10, V12, prescription isodose volume, BOT, PCI, and selectivity. The V10 (and V12) was significantly lower with the IP (p < 0.001). These volumes were 8.69 cm3 ± 11.39 and 5.47 cm3 ± 7.03, respectively, for the FP and 7.57 cm3 ± 9.44 and 4.78 cm3 ± 5.86 for the IP. Only the coverage was significantly lower with the IP (−2.3%, p < 0.001), but both selectivity (+17%) and PCI (+15%) were significantly higher with the IP than FP (p < 0.001). Conclusion: This IP demonstrated its capacity to generate multi-target plans rapidly, with a dose to the brain (V10) and BOT noninferior to the one of a human expert planner. These results would benefit from confirmation in a larger prospective series.