ESTRO 2021 Abstract Book

S267

ESTRO 2021

during gating would change the target coverage D95% with a median deviation from the static plans of -2.4% (- 13.3% to 6.4%). Applying the same treatment plans without gating would have resulted in a deviation of -10.3% (-41.7% to 6.2%). Similarly, the target mean dose Dmean deviated from the static plans by -1.1% (-4.6% to - 0.3%) with gating and -5.4% (-27.2% to -0.4%) without gating. The Dmean in the 2cm-ring showed median deviations of 0.0% (-1.2% to 4.9%) with gating and -1.1% (-10.2% to 1.6%) without gating. The ring Dmean showed decreasing values with increasing motion, but less pronounced as for the target doses, while the maximum dose in the ring D2% did not show any relevant correlation, but rather patient specific behavior (Spearman rank correlation: p<0.001 for all parameters, rho=-0.80,-0.80, -0.65, 0.27 for target Dmean, D95%, ring Dmean and D2%, respectively). In the figure, the percental differences of the dosimetric results with and without gating (top) and their correlations to RMSE (bottom) are visualized.

Conclusion The MRIdian system allows gating based on real-time imaging of the target anatomy and minimizes residual motion to 3 mm. The implemented dose reconstruction workflow allowed patient-individual quality assurance of the dosimetric accuracy of MR-guided beam gating. OC-0363 Evaluation of how well a PCA model represents anatomical variations during H&N radiation treatment J. Robbins 1 , R. Argota-Perez 2 , A. Green 1 , M. van Herk 1 , S. Korreman 2 , E. Vasquez Osorio 1 1 University of Manchester, Division of Cancer Sciences, Manchester, United Kingdom; 2 Aarhus University Hospital, Department of Oncology, Aarhus, Denmark Purpose or Objective Throughout a radiation treatment course, patients’ internal anatomy changes. One method commonly used to model and predict these anatomical changes is principal component analysis (PCA). The quality of these models must be evaluated to assess whether they can accurately predict unseen changes. In this study, we present a method to evaluate how well PCA models created early in the treatment of head and neck (H&N) patients can predict changes later in the treatment, highlighting any potential regions where the changes seen in the patients are not captured by the PCA model. We investigated two different scenarios: intra-patient and inter-patient models. Materials and Methods Intra-patient models were created using data from twenty H&N patients. An inter-patient model was created from a second dataset of ten patients plus eleven additional patients used for validation. The inputs for the PCA-models were deformation vector fields (DVFs) mapping CBCTs acquired for set-up correction and planning CTs (pCT), see Figure 1. For the inter-patient model, these DVFs were mapped to an average pCT of the training patients. The PCA models were created from the DVFs corresponding to the first week of treatment, i.e., 5 CBCTs for intra-patient models and 1 CBCT per patient for the inter-patient model. We evaluated how close the model could predict the variations seen in the rest of the treatment. For this, a predicted DVF was created by taking the inner product of each component from the PCA model and each ‘real’ DVF corresponding to the CBCTs after week 1 (minus the model’s mean). The magnitude of the difference between the predicted DVF and the ‘real’ DVF for each voxel ( M res ) described the spatial residual error of the PCA model.