ESTRO 38 Abstract book

S220 ESTRO 38

significantly (R^2=0.46) increased mean lung dose of up to 57% (range: 7.6%-11.6%, Fig. 2b). Generally, dosimetric impacts due to the variation of ITV volume definition are larger for inter- fractional than intra -fractional variability.

M. Krieger 1,2 , A. Giger 3 , D.C. Weber 1,4 , A.J. Lomax 1,2 , Y. Zhang 1 1 Paul Scherrer Institute, Centre for Proton Therapy, Villigen PSI, Switzerland ; 2 ETH Zurich, Department of Physics, Zurich, Switzerland ; 3 University of Basel, Center for medical Image Analysis and Navigation CIAN- Department of Biomedical Engineering, Basel, Switzerland ; 4 University Hospital Zurich, Department of Radiation Oncology, Zurich, Switzerland Purpose or Objective To investigate the consequences of respiratory motion pattern variability on ITV definition for treatment planning and 4D dosimetric evaluation. Material and Methods Synthetic 4DCT (4DCT(MRI)) data sets (Fig. 1b) have been generated by warping a single phase 3DCT using multiple motions extracted from 4DMRI datasets of a single subject (Boye et al 2013, Med.Phys. 40(6)). 20 such 4DCT(MRI) data sets have been generated from 20 respiratory cycles (128 phases) with large amplitude/period variations (Fig. 1a), which were then sorted as 4 sub-datasets (each of 5 cycles), each of which simulates motions that might occur on four different fractions. To evaluate the influence of intra - and inter -fractional motion variabilities from this data, two field PBS proton treatment plans were optimised on different geometric ITVs by considering several definitions (see Fig. 1c for details of each scenario). 4D dose calculations were applied to quantify the dosimetric effects by assuming different motion scenarios during dose delivery, using 9 volumetric rescans in order to minimise the influence of the interplay effect. Additionally, a ‘conservative’ ITV was calculated as a reference by considering all cycles together. Planning was performed on phase-averaged 4DCT(MRI) data sets, but using maximum intensity projections within the ITV. The resulting 4D dose distributions were analysed in terms of CTV coverage and homogeneity (V95%, D5-D95%) as well as mean lung dose. Linear regression was used to extract trends between relative ITV volumes and the aforementioned dose parameters.

Conclusion The ITV defined at treatment planning phase can vary up to ±20% depending on the motion cycle considered by a 4DCT and could become invalid when motions are considerably different during dose delivery, thus substantially compromising the quality of the dose delivered in a fraction. Thus, careful consideration of motion variability is required for 4D treatments at both planning and delivery phase. Acknowledgement This study was funded by Swiss National Science Foundation grant No 320030_163330/1. PV-0423 Fast automated IMRT sequencing using deep- learned dose from generative adversarial networks C. Kontaxis 1 , G. Bol 1 , J. Lagendijk 1 , B. Raaymakers 1 1 UMC Utrecht, Radiotherapy Department, Utrecht, The Netherlands Purpose or Objective In this work we utilize generative adversarial networks (GANs) to perform fast dose prediction along with a two- step IMRT optimization pipeline capable of fast automated plan optimization for prostate radiotherapy. Material and Methods A conditional GAN network was trained to predict the 3D clinical dose distributions from data of previously treated patients in our clinic. For 291 prostate patients the masks of the relevant optimization volumes, including bladder, rectum, femur heads, PTV and EBV were calculated. Then, for each 3 mm transversal PTV slice, a 256x256 image of the combined masks and background Hounsfield Units was extracted along with its corresponding slice from the clinical dose distribution of that patient (Figure 1). For each input mask slice i , a three-channel image was formed by placing slice i-1 into red, slice i into green and slice i+1 into blue channel respectively; in a caudal-to-cranial direction encoding the local 3D anatomical information into the 2D image. Then, the pix2pix framework was used to train the network with the data of 215 patients, totalling 5907 images. The resulting network was used to predict the dose distribution for the remaining independent 76 patients per slice. The 2D slices are then concatenated into the final predicted 3D dose distribution and fed into our previously developed Adaptive Sequencer (ASEQ) to perform IMRT optimization generating deliverable plans targeting that input dose. ASEQ is an IMRT optimizer which uses a per-voxel dose prescription along with minimum/maximum weights to penalize under- and overdosage. For automated planning a generic weights set is selected following the clinical structure priority:

Results Based on 4DMRI extracted motions, mean ITV size (29.4cc) was 156% of the static CTV (18.8cc), with large variations of ITV size of up to ±20% (range: 23.3-34.4cc) observed depending on the considered motion cycle during treatment planning (see Fig. 2a). Coverage (range: 82.7%- 99.7%) and homogeneity (range: 7.7%-20.0%) for all 4D plan scenarios (Fig. 2c-d) show a clear trend of improved coverage and homogeneity for plans optimised on the larger ITVs and encountering smaller motions during delivery (R^2=0.39 for D5-D95%, 0.28 for V95%). Moreover, considering the conservative ITV, the volume is between 16% and 43% larger than any individual ITV and induced a