|
内容記述 |
Purpose/Objective(s)While 4DCT is standard for proton therapy planning, its use is typically limited to treatment planning CT imaging. However, respiratory motion can change during treatment, and repeat 4D imaging is not routinely available for IGRT verification. This study evaluates whether synthetic 4DCT, generated from a single 3DCT using a deep neural network, can serve as a surrogate for motion-adaptive PBS proton planning. By demonstrating dosimetric accuracy, this approach may enable motion assessment using pre-treatment diagnostic scans and IGRT images.Materials/MethodsA deep neural network (DNN) was trained using 3DCT and 4DCT datasets from 436 patients (2,420 scans). The model predicts deformation vector fields (DVFs) from a single 3DCT, generating synthetic 4DCT images representing different respiratory phases. Training used a multitask learning approach to minimize errors between predicted and ground-truth DVFs from deformable image registration. To assess model generalizability, the DNN was tested on free-breathing 3DCT and ground-truth 4DCT datasets acquired from an independent institution not included in the training data. Synthetic 4DCT images were used to create maximum intensity projections (Synth-MIPs), which were used for PBS proton planning. Plans created with Synth-MIPs were compared to those optimized on Actual-MIPs from measured 4DCT. Both plans were robustly optimized using identical treatment geometry, objectives, and constraints.ResultsSynthetic 10-phase 4DCT images were successfully generated for five target volumes. Volumes for the Actual-MIPs were 13.2, 46.5, 48.6, 225.1, and 313.5 cm³, while the corresponding Synth-MIP volumes were 18.0, 41.2, 48.0, 222.6, and 326.5 cm³. PBS proton plans were created using Synth-MIPs and then evaluated on the Actual-MIP to assess dosimetric impact. This evaluation simulated the dose that would have been delivered had planning relied solely on synthetic motion data. ITV D95% coverage for these cases using 3-mm robust scenarios was 100.0%, 79.0%, 100.2%, 92.5%, and 91.8%, respectively. One target (79.0% ITV D95%) exhibited significant underdosage. This case was a lung retreatment and the difference in predicted target trajectory may be attributed to differences associated with the previous treatment. Additional strategies, such as incorporating a second 3D dataset or adding measured respiration traces may improve predictive accuracy and mitigate dosimetric discrepancies.ConclusionThis study demonstrates the feasibility of generating synthetic 4DCT from single 3DCT datasets using a deep neural network. Dosimetric evaluation showed strong agreement between synthetic and actual 4DCT-based plans, with one target exhibiting underdosage due to motion estimation limitations. Synthetic 4DCT may support motion assessment in settings where updated 4D imaging is unavailable. Refinements in motion modeling techniques could further enhance clinical applicability. |
