Author: Wenchao Cao, Yingxuan Chen, Haisong Liu, Wenyin Shi, Wentao Wang, Lydia J. Wilson, Zhenghao Xiao π¨βπ¬
Affiliation: Thomas Jefferson University π
Purpose: Multi-target stereotactic radiosurgery (SRS) planning poses challenges due to complex geometries, small target volumes, and steep dose gradients. Achieving a balance between target coverage and brain sparing requires specialized expertise. We propose a novel deep learning model, SRS Planning AI Dose Evaluation (SPADE), based on a conditional Generative Adversarial Network (cGAN), to predict optimal brain dose distributions and facilitate SRS plan evaluation.
Methods: The SPADE model was trained on data from 32 patients (328 target lesions) using target masks with prescribed doses (14β24 Gy) and a 1 cm exponential falloff. High-quality Gamma Knife dose distributions served as ground truth To enhance prediction accuracy, custom loss functions were implemented, including windowed mean absolute error (MAE), Sobel gradient, and Fourier loss. Model performance was evaluated on a test set of 30 patients (217 target lesions) using ablation studies to assess the contributions of each loss function. SPADEβs performance was also compared to a U-Net baseline. Evaluation metrics included overall MAE and root-mean-square-error (RMSE) for brain dose metrics: V12Gy, V8Gy, and V4Gy, representing brain volumes receiving at least 12 Gy, 8 Gy, and 4 Gy, respectively.
Results: Ablation studies demonstrated that incorporating custom loss functions reduced the test set MAE from 0.65 to 0.38. SPADE achieved an overall MAE of 0.83 Β± 0.22 Gy in dose prediction. Compared to the U-Net baseline, SPADE showed significant reductions in RMSE for V12Gy (0.36 vs. 0.55 cc), V8Gy (1.00 vs. 1.35 cc), and V4Gy (2.24 vs. 3.18 cc).
Conclusion: SPADE provides a reliable solution for generating benchmark dose distributions in multi-target SRS. Its accurate prediction of dose gradients in high-falloff regions enables robust quantitative evaluation of brain sparing. SPADE represents a valuable tool for treatment plan optimization and quality assurance in clinical practice.