ESTRO 38 Abstract book

S269 ESTRO 38

a per patient basis, where the first 200 frames were used for training and the last 100 frames for testing. The training time was 45 minutes using a nVidia Titan Xp GPU. For evaluation, motion fields of 1) the predicted frames (n+1)’ and 2) the most current frames (n) were calculated with respect to the ground truth frames (n+1). The motion fields were then used to calculate per-voxel displacement curves and RMSE maps, with and without prediction. Results The RMSE maps show that the error with prediction is substantially lower than without prediction in almost all areas of the image (figure 1). Displacement curves at points around the tumor (figure 2) show that phase and baseline of the predicted signal correspond with the ground truth, reducing the RMSE from 2.4 mm with no prediction to 1.1 mm with prediction, in this patient. Averaged over all patients, the RMSE was reduced from 1.47 mm to 0.74 mm, an average reduction factor of 2.0.

contours and -0.07±0.22% inside the high dose region (dose >90%). Dental artefacts obscuring the CT, could be circumvented in the sCT by the CNN-based approach in combination with TSE MRI sequence that typically is less prone to susceptibility artefacts (fig. 2).

Conclusion The 3D patch-based CNN generated sCTs of the H&N region that were dosimetrically accurate. The sCT were generated based on T2W TSE images already used for tumor/OAR contouring and thus no extra scan time was added. Moreover, for H&N cancers, the use of TSE as input for sCT generation has as particular advantage that it is less affected by dental artefacts compared to commonly used gradient echo sequences. OC-0516 Whole-frame 2D cineMR prediction using deep neural networks P. Borman 1 , L. Kerkmeijer 1 , B. Raaymakers 1 , M. Glitzner 1 1 UMC Utrecht, Radiotherapy, Utrecht, The Netherlands Purpose or Objective Accurate target localization during radiotherapy treatments is becoming increasingly feasible with the recent clinical introduction of MR-guided radiotherapy systems such as the Elekta Unity and ViewRay MRIdian. These systems will facilitate real-time treatment adaptation based on continuous MR imaging, to correct for e.g. respiratory motion. It is important to minimize the latency of this control loop, which is largely determined by the MR imaging due to its relatively long acquisition time [1]. The control loop latency may be reduced by predicting the next state based on previous states, a concept that has been previously applied to 1D respiratory signals [2]. In this work we present full frame, image- based, prediction results using the readily available deep learning library pix2pix [3]. Material and Methods MR images were acquired on a 1.5T Philips Ingenia system in five renal cell carcinoma patients, with IRB approval. A 2D balanced gradient echo sequence was used consisting of 300 sagittal frames, with a frame rate of 2 Hz. A TensorFlow implementation of pix2pix was used to train a model that predicts a frame based on three previous frames. For each frame (n), the input images consisted of frames (n-2), (n-1), and (n), while the target image was set to frame (n+1). The model was trained and tested on

Conclusion A deep learning based prediction model is indeed able to learn a patients breathing pattern and predict a full frame 500 ms ahead. This may help to reduce the influence of imaging latency, which is important for real-time MR-

guided treatment applications. [1] Borman et al. PMB 63(15) [2] Sun et al. PMB 62(17) [3] Isola et al. arXiv:1611.07004

OC-0517 Automatic tumor delineation in rectal cancer using functional MRI and machine learning F. Knuth 1 , A. Rosvoll Grøndahl 2 , T. Torheim 3,4 , A. Negård 5,6 , S.H. Holmedal 5 , K.M. Bakke 7,8 , S. Meltzer 7 , C. Futsæther 2 , K.R. Redalen 1,7 1 Norwegian University of Science and Technology, Department of Physics, Trondheim, Norway ; 2 Norwegian University of Life Sciences, Faculty of Science and Technology, Ås, Norway ; 3 University of Oslo, Department of Informatics, Oslo, Norway ; 4 Oslo University Hospital, Institute for Cancer Genetics and Informatics, Oslo, Norway; 5 Akershus University Hospital, Department of Radiology, Lørenskog, Norway; 6 University of Oslo, Institute of Clinical Medicine, Oslo, Norway ; 7 Akershus University Hospital, Department of Oncology, Lørenskog, Norway ; 8 University of Oslo, Department of Physics, Oslo, Norway Purpose or Objective Tumor delineation is a time- and labor-intensive procedure both for radiotherapy planning and for quantitative imaging biomarker purposes. In addition, it is