ESTRO 38 Abstract book

S969 ESTRO 38

XVMC for C1, collapsed cone convolution for the others. Each TPS calculated the dose for two plans (both with 6MV beams at DFA 95cm): a 10x10cm 2 beam and a dynamic multileaf collimator plan DMLC 10x10cm 2 field as a series of 10 rectangular strips in the lateral direction of 10 cm x 1 cm each. Each plan delivered 2Gy in one fraction, with a grid resolution of 2mm. TPSs calculated DVH for the PMK structure when its proximal edge was 1 cm to 10 cm to the field edge. Results are expressed as cGy/Gy. TPSs comparison is performed on the PMK maximum dose (D1cc) and mean dose (Dmean). An analysis is performed on the difference between these doses and the maximum discrepancies were evaluated.

EP-1791 Correcting dose distributions to the magnetic field of a high-field MR-Linac using deep learning R. Rozendaal 1 , W. Van den Wollenberg 1 , S. Van Kranen 1 , J. Sonke 1 Netherlands Cancer Institute, Department of Radiotherapy Physics, Amsterdam, The Netherlands Purpose or Objective The Elekta Unity MR-Linac integrates a high-field MR with a state of the art Linac. In order to exploit the potential of this device, the effect of the magnetic field must be taken into account when calculating dose. This is typically done using a Monte Carlo (MC) method, which is accurate but also computationally expensive. The associated computation time adds to the daily treatment time slots. The purpose of this study was to evaluate the potential of a convolutional neural network (CN)N to account for the presence of a magnetic field. Material and Methods The CNN architecture chosen was a 3D U-net [1] like with 3-fold downsampling. Input data consisted of beam specific 3D dose distributions calculated using Unity treatment planning system (Monaco, Elekta AB, Stockholm) with the magnetic field effects switched off, desired output data were the same dose distributions with the magnetic field on. In total 7 patients, each with one treatment plan, were included; training and validation was done using leave-one-out cross validation. Training samples consisted of a full 3D dose distribution composed of 9 beams and were generated on the fly. Each plan had 27 treatment beams available: 9 beams at the (fixed) gantry angles used for plan optimization, and then each beam offset by +/- 10 degrees. By not using the anatomy of the patient as an input to the network, 18^9 possible dose distributions available for training: 6 (patients) * 3 (beams per gantry angle) for 9 gantry angles. The loss function was a weighted mean squared error, with higher weights for higher doses. The network was implemented using python/keras; training was done on a NVIDIA GTX- 1080 GPU. After training, the predicted dose distribution for validation was evaluated by inspecting parameters of the dose-volume histogram (DVH): D2 (near-maximum dose), D50 (median dose) and D98 (near-minimum dose) for the planned target volume (PTV); mean dose for two organs-at-risk (OARs), the rectum and the anal sphincter. Values were calculated for the baseline-situation (no magnetic field corrections) and the CNN-predicted dose distribution, both versus ground truth, MC with magnetic field on. Results Training converged typically after ~2000 iterations, taking ~900s. A single prediction of the magnetic-field corrected dose took 1.64 ±0.04 s. Table 1 shows average results, Figure 1 shows typical dose-volume histogram (DVH) curves. Average correspondence in the PTV is within 1%, correspondence in the OARs is at -3% and -2%. In general, the effect of the magnetic field is better modelled for higher dose regions.

Results Out of field doses calculated at 1 cm and 2 cm depth are slightly the same for all TPSs: dose differences are less than 1 cGy/Gy for D1cc and Dmean, for both plans (see fig1). Discrepancies between D1cc and Dmean calculated form TPSs were higher for the DMLC plan than the 10x10cm 2 field as shown in fig2. For the square beam and DMLC plan, we identified two different regions of behavior calculations, as shown in fig2: for the square beam differences in D1cc and Dmean are higher close to the beam edge (distances ≤ 3 cm), whereas for the DMLC plan the discrepancies are greater at distances ≥ 7 cm from the beam edge.in the TPS Conclusion While with the square beam all TPSs are in good accordance in the out of field dose calculation (maximum difference 1.6 cGy/Gy), in the DMLC plan the difference between TPSs is higher (up to 4.5 cGy/Gy). TPSs show different behavior in out of field dose for the square beam with respect to the DMLC plan, probably related to collimator scatter and head leakage modeling. All centers are currently working on measurements of such doses to better understand and evaluate TPS’s reliability in estimating dose to a PMK.