Author: Wenchao Cao, Yingxuan Chen, Haisong Liu, Richard A. Popple, Wenyin Shi, Rodney J. Sullivan, Wentao Wang, Lydia J. Wilson, Zhenghao Xiao 👨🔬
Affiliation: Thomas Jefferson University, The University of Alabama at Birmingham, University of Alabama at Birmingham 🌍
Purpose: Objectively evaluating brain sparing as a plan-quality indicator for patients receiving stereotactic radiosurgery (SRS) to multiple metastases (multi-met) is complicated by variability in target geometry and prescription dose. This study introduces the Brain Sparing Index (BSI), an AI-powered tool that evaluates target-specific brain sparing to guide multi-met SRS optimization.
Methods: BSI compares the planned dose to an AI-predicted benchmark dose with ideal brain sparing. Within a 2-cm target ring, the area under the curve of the dose-volume histogram (AUC-DVH) is calculated between 50-80% Rx. For each target, BSI is the AUC-DVH ratio of the benchmark dose to the planned dose. Higher BSI indicates better dose sparing to the normal brain, with BSI = 1 indicating the same brain sparing as the benchmark.
To predict benchmark doses for BSI calculation, SPADE (SRS Planning AI Dose Evaluation), a conditional generative adversarial network, was developed using 62 patients, split into 32 for training and 30 for testing. Manual benchmark plans (high-quality Gamma Knife plans) achieving optimal brain sparing were created to train SPADE. Dynamic conformal arc (DCA) and HyperArc (HA) plans were created for the test set as clinically relevant plans for BSI calculation.
Results: SPADE accurately predicted the benchmark dose within target rings among 217 targets in the test set (Rx: 14-24Gy): voxel-wise absolute value error was 0.83 ± 0.22 Gy (mean ± standard deviation); V12Gy absolute value error was 0.18 ± 0.31 cc. Using AI-predicted dose as the benchmark, the median BSI [25th-75th percentiles] was 1.01 [0.88-1.16] for manual benchmark plans, 0.62 [0.52-0.72] for DCA plans, and 0.56 [0.47-0.65] for HA plans.
Conclusion: Powered by patient-specific benchmark doses from SPADE, the BSI can facilitate efficient evaluation of brain sparing by quantifying deviations from optimal dose and guide multi-met SRS optimization.