ESTRO 38 Abstract book

S268 ESTRO 38

The accuracy of both methods was determined with a modified distance discordance metric[2] (DDM): a deformation vector field (DVF) between a planning CT (pCT) and a repeat CT (rCT) was constructed by concatenating DVFs between the pCT and a CBCT and the CBCT and the rCT. With 35 CBCTs per patient, the distribution of the DVF end points was determined (per pCT voxel as the vector length of the standard deviation per direction). Finally the accuracy for relevant structures was quantified. The DDM was collected over a held-out test set of 5 patients. Results Training typically converged within three epochs. Regularization weights in the range 0.5-2 showed a shallow optimum and resulted in an absence of folding effects and unrealistic micro deformations. Applying trained models to scan-pairs took ~10 sec, vs minutes with our bSpline method. Figure 1 shows an example DDM color overlay. VoxelMorph generally showed accurate mapping, but failed at occasional large deformations (weight loss, shoulders) and areas with sliding tissue (uvula, epiglottis) or (dis)appearing air gaps (nose, oral cavity). Figure 2 shows the distribution of DDM vector lengths of target areas and various organs. Both deformation methods showed excellent mapping-accuracy: 50% of voxels within the patient- external lay within 0.5mm, 90% within 1.5mm. Outliers within the external were found to correspond with above mentioned failures. Specified per structure, bSplines DIR was more accurate than VoxelMorph. [1] Balakrishnan G, CVPR, 2018 [2] Saleh ZH, PMB, 2014 Conclusion CNN based DIR resulted in fast and accurate head & neck CBCT-to-CT mappings, creating opportunities for on the fly DIR for online plan adaptation. Accuracy was acceptable but less than our clinically applied bSpline DIR and failed at occasional large deformations. Diversifying the training with large deformations and other tumor sites may improve accuracy and robustness, and prevent failures.

OC-0515 Synthetic CT generation for Head and Neck radiotherapy by a 3D convolutional neural network A. Dinkla 1 , M. Florkow 2 , M. Maspero 1 , M. Savenije 1 , F. Zijlstra 2 , P. Doornaert 1 , M. Van Stralen 2 , M. Philippens 1 , P. Seevinck 2 , N. Van den Berg 1 1 UMC Utrecht, Department of Radiation Oncology, Utrecht, The Netherlands; 2 UMC Utrecht, Centre for Image Sciences, Utrecht, The Netherlands Purpose or Objective Synthetic CT generation is required for MR-only simulation workflow where CT-MRI registration can be avoided. More recently, also sCTs are becoming desirable for MRI-guided Radiotherapy where planning is performed based on daily MR images. In the Head and Neck region (H&N), atlas-based methods have been adopted. However, their robustness is limited as in H&N abnormal patient anatomies can occur due to large tumors or surgical excision. Therefore, here a patch- based deep learning method was chosen to improve robustness. In particular, we used a 3D patch-based convolutional neural network (CNN) to generate sCTs based on T2-weighted Turbo Spin echo (TSE) images and evaluated its image and dosimetric performance. Material and Methods We conducted a retrospective study on 34 patients with Head and Neck cancer who underwent CT (Philips Brilliance Big Bore) and MR imaging (3T Philips Ingenia) for radiotherapy simulation. To generate the sCTs, a large field-of-view (FOV) transverse T2-w TSE mDixon MRI, originally used for tumor/OAR contouring, was selected from the clinical protocol. 83 transverse slices with 3 mm thickness were acquired with a FOV of 45x45 cm2 and 0.94x0.94 mm2 resolution in 5min24s and readout bandwidth of 876 Hz/px. Cases with severe image artefacts from dental implants (CT) or motion (MRI) were excluded from the training. To align images for training and evaluation, CT scans were non-rigidly registered (CTreg) to the in-phase MR images (Elastix 4.7) and all images were resampled to 1x1x1 mm3 isotropic resolution. The CNN was based on a U-net architecture and consisted of 14 layers with 3x3x3 filters. Patches of 48x48x48 were randomly extracted and fed into the training. sCTs were created for all patients using three-fold cross validation. The CT-based treatment plan was recalculated on sCT using Monaco TPS (Elekta). Results sCT generation took 4 min. on a single GPU. The patch- based approach allowed proper sCT generation for non- standard anatomies (fig. 1). Mean absolute error (MAE) over the patient population of the sCT within the intersection of body contours was 75±9 HU, and the mean error (ME) was 9±11 HU. Dice scores of the air (<-200HU) and bone (>250HU) masks (CTreg vs sCT) were 0.79±0.08 and 0.70±0.07 respectively. Dosimetric analysis showed mean differences of -0.03±0.05% for dose within the body