MRI is increasingly integrated into the radiotherapy treatment workflow because of high soft tissue contrast and tumor visibility. To fully exploit the benefits of MRI a stand-alone MRI workflow for radiotherapy is desirable. However, CT remains the standard since it provides electron density information, which is a requirement for treatment planning. Therefore the generation of pCT from MR images is investigated e.g. by tissue classification. In head and neck separating air and bone voxels is especially challenging because of neighboring location and fast signal decay of bone.

In this work the goal is to address geometrical accuracy of different UTE based sequences, its effect on pCT generation and the possibility to accelerate acquisition by undersampling.

Image acquisition: Two triple echo UTE sequences were tested: a 3D in-plane radial k-space acquisition “stack-of-stars” (SOS) sequence (TR/TE/ΔTE (ms) 6.9/0.14/1.1; FOV 232x232x256 mm^3, FA 10°, 1.5x1.5x2.0mm3) and a 3D radial “kooshball” (KOOSH) sequence (TR/TE/ΔTE (ms) 6.5/0.14/1.1; FOV (232x232x232 mm3), FA 10°, 1.5x1.5x1.5 mm3) on a 3T scanner (Ingenia, Philips, Best, Netherlands) with a 8-channel head coil. For both sequences radial undersampling up to 60% was tested in order to decrease scanning time. Each sequence was tested on two volunteers and compared for image acquisition convenience. pCT results are compared to the corresponding result from undersampled data.

Image-processing: An image-processing pipeline for synthesizing a pCT based on a triple echo UTE sequence was established (Fig.1) based on a single scan. First, a bone-enhanced image is derived, using the first and third echo magnitude images 2(M1+M3)/(M1+M3) [1](Fig.2a). The third echo was used to derive both a body mask for segmentation and a water-only pCT. A two-point Dixon decomposition [2] identified soft tissue voxels in the bone-enhanced image (Fig 2b,c). Air voxels within the patient are classified by Otsu-thresholding and morphological filtering on M1 (Fig.2d,e). A threshold operation was applied to isolate bone voxels. Finally body, bone and air-inside mask are assigned bulk density according to literature values [3]. Image processing was performed using IPython and the Insight Toolkit (SimpleITK) [4,5].

Geometric accuracy: For verifying geometric accuracy, the sequences were tested on a realistic agar head phantom comparing body outline differences for different echoes and parameter settings to a high resolution CT image.

Fig.2f shows the final bone segmentation resulting in a bone mask visible in Fig.2i. Fig.2g presents the resulting pCT and Fig.2h the Digital Reconstructed Radiography (DRR) image for position verification. pCT generation takes 1:05 min (8 core CPU computer), the imaging time employing radial undersampling is 3:01min. Fig.3 displays the results comparing the SOS and KOOSH sequence including the result for radial under sampling. Tissue classification is compared in percent to the body volume. Radial undersampling leads to a percentage difference in the order of <1% for the KOOSH sequence and to ≈ 4% difference in SOS bone and water classification for volunteer1 and <1% difference for volunteer2. Geometrical accuracy, tested on the agar-phantom showed a match in body outline of CT to M1(M2, M3) to 95.0%(98.2%, 98.8%). Comparing body outline for the same echo time but different trajectory delays (1,2,10,-10 μs) resulted in a difference of 1%.

Image acquisition and image processing time in total is in the order of 4 min if radial undersampling is used, which is within the requirements for MR-linac applications [6]. Comparing the feasibility of the sequences, the SOS gives more freedom and flexibility in tradeoffs for resolution and scanning time, especially for large FOV including shoulders.

The image-processing pipeline has to be tested on an increased number of datasets to ensure robustness. It can be potentially improved employing a conversion of HU based on intensity levels instead of a threshold on the final bone-enhanced image in combination with bulk assignment [7]. Also air segmentation robustness has to be validated and potentially improved [8]. An open problem are dental implants and metal wire artifacts, which were observed both datasets. It has to be seen if manual bulk assignment in the artifact area is clinically feasible and dosimetrically accurate.

The results suggest that pCT generation for Head and Neck within clinical acceptable time is feasible. As a next step we aim to verify the tissue classification with CT data and compare dose outcome of a clinical plan on pCT and CT. The workflow will be adapted and tested for realistic radiotherapy treatment, which involves imaging in immobilization mask.