Knowledge-based VMAT planning by using the dosimetric features for cervical cancer

Abstract

Significant dose variations among cervical cancer patients with similar geometric characteristics may limit the effectiveness of knowledge-based planning (KBP) techniques that rely solely on geometric features for volumetric modulated arc therapy (VMAT) planning. To address this issue, this study adapts methodologies from prior research and proposes integrating dosimetric features into the VMAT plan generation process for cervical cancer. Seventy previously treated cervical cancer patients were included in the training cohort, while fifteen additional cases were selected as the testing set. The study employed two features—the distance-to-target histogram (DTH) and the unit-fluence-dose volume histogram (UFDVH)—to predict dose-volume histograms (DVHs). The generalized regression neural network (GRNN) was used as the DVH prediction model. Based on these predictions, DTH-based and UFDVH-based plans were generated, each adhering to their respective dose-volume constraints. The resulting plans were then evaluated using the paired t-tests and Wilcoxon Signed Rank test statistical analysis. The target dose coverage was comparable across all plans. Both the DTH plan and UFDVH plan performed better than manual plan in sparing some OARs. While UFDVH could deliver a lower dose volume at left femoral head (V20(%): 82.50 versus 88.39, p < 0.05; V50(%): 28.38 versus 31.32, p < 0.05; V60(%): 16.18 versus 18.55, p < 0.05) than DTH method without reducing target dose coverage. Both the two KBP methods could improve planning efficiency when compared with manual plan. This work diversifies the KBP method application by using the dosimetric features UFDVH. The results indicated that the dosimetric features can be successfully used for knowledge-based planning in cervical cancer VMAT plan generation and improves the planning efficiency.