ESTRO 2024 - Abstract Book

S4089

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2024

fine-tuning the BM, personalized single-patient models were trained from scratch with either the planning image (PS noBM F0 ), or with the planning and first four fraction MRIs (PS noBM F0-4 ). Since training without BM involved only up to five images, 2D models were implemented instead of 3D ones. This reduced network complexity and increased the number of training examples by treating each axial slice independently. Fraction data from 23 patients was used to determine optimal hyperparameters for all PS methods. Finally, each technique was tested on the remaining 21 patients, employing the pre-determined hyperparameters and a fixed number of epochs. All methods shared the same test set, i.e., the fifth fraction MRIs from 21 patients. Evaluation criteria included Dice similarity coefficient (DSC), average (HD avg ) and the 95 th percentile (HD 95 ) Hausdorff distance. A radiation oncologist working clinically in MRgRT assessed the quality of predicted OAR contours for plan adaptation. The grades, ranging from 0 to 4, corresponded to the following statements: 0-no clinically relevant corrections possible, 1-ready to use, 2-minor corrections required, 3-major corrections required, and 4-unusable. In this analysis the BM, PS BM F0 , and PS noBM F0 were included as methods usable already from the first fraction.

Results:

F0-4 followed by PS

F0

Figure 1 displays the geometric performance of the investigated methods on the test set. PS BM

BM

networks demonstrated the best results across all organs.