Purpose/Objective: For RT based on MRI only, a promising approach is to obtain a substitute CT scan from the MR images (a so-called pseudo CT, pCT) using a Gaussian mixture regression (GMR) model. The GMR model has previously been investigated on 3T MR images using a dual ultra-short echo time (dUTE) sequence and was shown to give sufficient information for training the GMR model. The dUTE sequence provides contrast between bone and tissue using dual echo times but at 1 T, chemical shift

more »

... ts at the second echo time may cause voxels containing water and fat to behave like bone. The multiecho-Dixon (mDixon) MR sequence provides contrast between water and fat which could potentially remove this problem and provide valuable information for the GMR model. In this study, we investigate the robustness of the GMR model on predicting pCT scans from dUTE MR images of a 1 T scanner and how adding an mDixon sequence affects the generated pCT. Materials and Methods: Head scans of 2 patients fixated for whole brain RT were acquired on a 1 T open MR scanner with flex coils. dUTE sequences were obtained at flip angles 10 and 25 degrees, respectively. Echo-and repetition times TE1/TE2/TR were 0.09/3.5 /7.1 ms with a voxel resolution of 1x1x1 mm and a 256 mm FOV. The mDixon was acquired with TE1/dTE/TR equal to 6.9/3.5/16 ms, a voxel resolution of 1x1x1.5 mm and a 250.5 mm FOV. CT head scans were acquired with a voxel resolution of 0.6x0.6x2 mm and a 220 mm FOV. The CT was registered to the high angle TE1 UTE using a mutual information algorithm and all MR scans were internally registered. All scans were resliced to the dUTE resolution and cropped to the smallest FOV. The MR images were low-and high-pass filtered creating two new images per filtered image. The MR images, their filtered counterparts and the CT image were considered as random variables and the voxel intensities a sample from their underlying distribution. A GMR model was initialized with 20 centers using kmeans clustering and an EM algorithm was used to train the model on the data from one of the patients.The model was then applied on the other patient to generate the pCT. A model using only the dUTE images and one adding the mDixon images were trained.A comparison using the real CT to calculate the mean absolute prediction error (MAPD) of the pCT in bins of 20 HU was carried out. Results: The pCTs of one patient using the extended model is shown in the figure. Qualitatively (upper images) and quantitatively (lower graph), the results are similar to those previously reported for 3T using dUTE only. A reduction in MAPD can be observed in the bone region(>500 HU) by adding mDixon to the model. Conclusions: The robustness of a GMR model on 1T MR images was demonstrated. The model was further expanded with an mDixon sequence which reduced the prediction error of predicted CT values >500 HU. Although a study based on larger amounts of data should be carried out, there is an indication that the mDixon sequence improves CT prediction from dUTE MR images. PD-0594 The introduction of simultaneous PET/MRI to radiotherapy planning