ESTRO 37 Abstract book

S150

ESTRO 37

assessing whether a cGAN can generate sCT suitable for treatment planning on prostate cancer patients. Material and Methods A study was conducted on 59 patients who underwent prostate IMRT for which CT (Brilliance CT Big Bore, Philips) as well as MR (3T Ingenia Omega HP, Philips) scans were acquired for simulation purposes on the same day and in RT position. To generate the sCT images, dual gradient-echo, RF spoiled, 3D T1w MR images were acquired with 1x1x2.5 mm3 resolution, TR/TE2/TE1=3.9/2.5/1.2ms. Dixon reconstruction was performed producing water (W), fat (F) and in-phase (IP) images. A cGAN called “pix2pix” (P. Isola et al, 2016, arXiv) was used in this experiment. Before training, CT images were rigidly registered to MR images and MR images were normalised (Figure 1a). Within pix2pix, the images were scaled to 8-bit grayscale values and resampled to 256x256-pixel slices (Figure 1b). Training of pix2pix was performed on 32 patients in 2D transverse slices using 200 epochs. To guarantee consistent air pockets on the CT and MR images during training, internal air pockets as detected on MR images were copied to CT (Figure 1c). The trained cGAN was applied to the remaining 27 patients (test set) producing sCT images. Image evaluation was performed on the sCT using mean absolute error (MAE) as compared to CT. Dose recalculation of clinical 5-beam 10 MV IMRT plans with a prescribed dose of 35x2.0 Gy to the target was performed for 15/27 patients on the CT and sCT images in Monaco (v 5.11.02, Elekta AB). Dose distributions were subsequently analysed through voxel-based dose differences and gamma analysis.

Image CT- sCT (HU)

Dosimetric CT-sCT (%)

Soft tissue

In PTV (%)

Total

Bone

>50% (%)

MAE (SD) ME (SD)

67.4 (11.2)

174.1 (29.2)

21.6 (2.8) 0.58 (0.2)

0.78 (0.4) -0.13 (0.4)

-1.6 (3.3)

12.5 (9.0) 74.6

-0.08 (0.3)

(41.3)

Conclusion The CNN trained in this study was able to generate a sCT volume in 1 min. The errors between original CT and sCTs were small and calculated dose distributions showed high accuracy. Besides being suitable for MR-guided RT which requires online sCT generation, this is a first step in achieving a fast MR-only workflow (‘one stop shop’) to decrease treatment time and burden to the patient. OC-0294 MR-based synthetic CT with conditional Generative Adversarial Network for prostate RT planning M.H.F. Savenije 1,2 , M. Maspero 1,2,3 , A.M. Dinkla 1,2 , P.R. Seevinck 2,3 , C.A.T. Van den Berg 1,2 1 UMC Utrecht, Department of Radiation Oncology, Utrecht, The Netherlands 2 UMC Utrecht, Center for Image Sciences, Utrecht, The Netherlands 3 UMC Utrecht, Image Science Institute, Utrecht, The Netherlands Purpose or Objective To enable MR-only planning and accurate MR-based dose calculations, so-called synthetic CT (sCT) images need to be generated. Recently, conditional Generative Adversarial Networks (cGANs) have been proposed as a general-purpose solution to image-to-image translation problems. By interpreting the generation of sCT images as an image-to-image problem, this work aims at

Results In total, 3495 slices were used for training, requiring about 11 hrs on a GPU Tesla P100 (NVIDIA). Applying the trained cGAN to a single patient volume (Figure 2) required about 5.6s. Image Evaluation: On average, an MAE of 65±10HU (±1σ, range: 50-96HU) was obtained in the intersection of the body contours between CT and sCT. When air pockets were also copied to the CT in the test set, the MAE reduced to 60±6HU (range: 48-71HU). Dose Evaluation: On average (Table 1), a dose difference below 1.1% was obtained using a 50% dose threshold of the prescribed dose. A mean gamma pass rate of 96±4% was obtained.