Enhancing Direct VMAT Plan Generation for Cervical Cancer Via 3D Dose-Domain Supervision
Abstract
Purpose
Volumetric modulated arc therapy (VMAT) planning is complex and time-consuming. This study proposes a direct VMAT plan generation method for cervical cancer, enhanced by a physics-guided 3D dose-domain supervision strategy to ensure dosimetric quality.
Methods
A total of 633 cervical cancer VMAT plans were retrospectively collected and split into training (445), validation (128), and testing (60) cohorts. Target and OAR masks, encoded with prescription doses and organ indices, were projected into two-dimensional beam's-eye-view (BEV) space for each control point as inputs. A convolutional encoder was employed to encode each BEV projection, followed by an attention mechanism to model temporal dependencies across control points. Finally, three distinct decoder paths directly output the MLC leaf positions, Jaw positions, and Monitor Units (MU). The network was trained using a two-stage strategy: the first stage utilized original plan parameters for direct supervision, while the second stage applied 3D dose-domain supervision using 3D fluence volumes and 3D dose distributions to optimize the dosimetric performance.
Results
The directly generated plan parameters demonstrated high consistency with clinical plans. The second-stage 3D dose-domain supervision significantly improved dosimetric performance. The D95 of PTV60 and PTV45 was refined from 59.43 ± 6.82 Gy and 46.76 ± 6.03 Gy to 59.34 ± 1.75 Gy and 45.58 ± 0.65 Gy, respectively, with markedly reduced variability, indicating improved stability in prescription dose coverage. Additionally, maximum doses to the spinal cord, bladder, rectum, and bilateral femoral heads were reduced to clinically acceptable ranges. DVH analysis further confirmed that the dose-domain optimization significantly enhanced target coverage while improving OAR sparing. The mean computation time is about 290 ms.
Conclusion
This study presents a direct automated VMAT plan generation method for cervical cancer, distinctively enhanced by 3D dose-domain supervision. The generated plans are clinically acceptable or require only minor modifications, thereby significantly improving planning efficiency.