Author: Chih-Wei Chang, Runyu Jiang, Mark Korpics, Yuan Shao, Aranee Sivananthan, Zhen Tian, Ralph Weichselbaum, Xiaofeng Yang, Aubrey Zhang, Xiaoman Zhang 👨🔬
Affiliation: Department of Radiation & Cellular Oncology, University of Chicago, University of Chicago, Department of Physics, University of Chicago, Emory University, Department of Radiation Oncology and Winship Cancer Institute, Emory University, School of Public Health, University of Illinois Chicago 🌍
Purpose: Gamma Knife (GK) plan quality can vary significantly among planners, even for cases handled by the same planner. Although plan quality metrics such as coverage, selectivity, and gradient index (GI) are commonly used to evaluate GK plans, the best achievable quality is essentially determined by patient-specific geometry. Comparable metrics across cases do not necessarily indicate the same quality level. This study proposes a deep learning-based method to predict the achievable GK plan quality based on patient-specific geometry in patients with brain metastases (BM) for quality control purpose.
Methods: We employ a hierarchically densely connected UNet architecture to predict 3D dose distributions in order to predict the achievable plan quality metrics, including coverage, selectivity, GI, and intermediate dose spillage (CI50). The network was trained on individual BMs using 3D mask data, where voxel values represent the patient skull and BM volume. To enhance the accuracy of predicted quality metrics, we design dice similarity coefficient losses for the 100% and 50% isodose lines and incorporate them into the conventional mean squared error (MSE) loss.
Results: A retrospective study was performed on 175 patients with 463 BMs included in total, using ten-fold cross-validation. Compared to the MSE loss baseline, our approach achieved smaller values of the reduced χ2 in predicting selectivity (0.02 vs. 0.04), GI (0.29 vs. 0.33), and CI50 (0.73 VS. 0.92), while maintaining comparable coverage prediction. The improvements were particularly pronounced for smaller BMs, with the reduced χ2 values improved from 0.08 to 0.04 for selectivity, 0.86 to 0.72 for GI, and 2.24 to 1.84 for CI50.
Conclusion: These results demonstrate the feasibility and efficacy of the proposed method for GK plan quality prediction. Beyond quality control, this method may also assist in automated treatment planning by guiding personalized objective function weights or providing reference quality for intermediate plan evaluation.