Abstract Book

S349

ESTRO 37

SP-0660 Different methods of creating pseudo-CT images J. Jonsson 1 1 University Umea Norrlands Universitetssjukhus, Radiation Sciences, Umea, Sweden Abstract text Ever since the introduction of MR guided target definition for radiotherapy, the notion of MR-only based radiotherapy has been present. Since the target delineation is based on MR images, MR-only based radiotherapy would be favorable to simplify the workflow, as well as to reduce the risk of systematic errors being introduced into the treatment from image registration between the MR and planning CT. A major obstacle to overcome before such a workflow can become reality, is to be able to perform heterogeneity corrected dose calculations without the CT image. While the CT directly reflects the attenuation properties of the patient, the MR image correlates with the proton density and the magnetic relaxation of the imaged tissue. The typical example of the problem is the signal ambiguity between bone and air – while they both appear dark in an MR image, their attenuation coefficients are vastly different. This has led many research groups to investigate methods of constructing attenuation maps, or “pseudo CTs”, from MR data. There has been a plethora of methods presented in the literature, ranging from the very simple approach of assigning the entire patient anatomy the attenuation property of water, to much more involved conversions. The different methods can be categorized into two main approaches; voxel based and atlas based conversions. The voxel based methods mainly use the image intensities in the MR images for conversions – the spatial location of the voxels plays little or no role in the assignment of attenuation properties. Conversely, the atlas based methods mainly depend on the spatial location of the voxels to decide what attenuation property should be assigned to the pCT. There are also hybrid methods, where both voxel and atlas based methods are employed to reach the final assignment of attenuation. Most methods presented in the literature produce clinically acceptable results for photon dose calculations. In general, the voxel based methods are specific to the sequences which the models are trained on, but are general with regards to the patient population. In contrast, atlas based methods are more insensitive to input image data but specific to the patient population. In this lecture, the different methods will be discussed and exemplified, providing typical results and potential difficulties with the different conversion strategies. Abstract text MRI is increasingly used in external beam radiotherapy planning for target and normal structure delineation because of superior soft tissue contrast. The interobserver variability in defining target is significantly reduced with MR as compared to CT for many disease sites. Recently, MR simulation platforms have been introduced, including flat table tops, external laser systems, and radiation therapy specific scanning protocols that are further enabling the use of MRI as the primary imaging modality for radiotherapy planning. MRI as a primary imaging modality for treatment planning is preferred based on several major advantages, including minimizing dosimetric errors introduced by mis- registration with the planning CT or changes in anatomy between the two scans, improving efficiency, and reducing redundant imaging and patient costs. However, SP-0661 Clinical implementation and clinical experience with MR only workflows N. Tyagi 1 1 MSKCC, Medical Physics, New York- NY, USA

and needs to be supplemented by proper sequence design (e.g. a high readout bandwidth). A second adaptation crucial for MR-only simulation is the possibility to scan the patient in treatment position. Both aspects have been acknowledged in the last years by MR vendors and they introduced wide bore MR systems and launched special MR-RT solutions such as flat table tops, QA phantoms+procedures, positioning lasers and special coil options. Much of the research towards MR-only has focused on different methodologies to generate socalled synthetic CT from MR images. A crucial aspect is the visualization of the bony anatomy with MRI. Due to the low proton spin density and very short T2*, cortical bone appears as a signal void on MRI. Simple segmentation of these signal voids to identify cortical bone on MR images will not be successful as inner air also appears a signal void. Dedicated MR sequences such as ultra short echo time (UTE) sequences to obtain signal from cortical bone have been applied with mixed success. Other methods have approached the problem as a quantitative MRI problem to convert quantitative MR images into synthetic CTs using signal-to-houndfields unit conversion models. Recently, various commercial solutions have become available that use rather standard MR sequences combined with clever image processing solutions to generate synthetic CT images. A very new technique, deep learning based synthetic CT generation is even more flexible in terms of requirements for image contrast. Deep learning methods like convolutional neural networks are able to classify bony or air voxels by a learning approach including local, contextual image information. With these new deep learning methods standard MR sequences that are primarily intended for delineation purposes can be utilized for synthetic CT generation. More general, deep learning methods offer great potential for other steps in the MR-only simulation process, e.g. for automatic contouring of organs-at-risk on MR images. A somewhat overlooked aspect of MR-only simulation is the need to generate reference images for position verification from MR images. The synthetic CT will suffice as a reference image for tumor sites where position verification is based on registration of kV or MV in-room images to a reference image. However, the situation is more challenging in the case of position verification for VMAT or IMRT prostate irradiation where implanted gold fiducials are used. Fiducials are easily localized on CT images due to their distinct local streaking artifacts. However, on MR images the appearance is less distinct and they manifest themselves as local signal voids in magnitude MR images. Correct manual classification of these signal voids as fiducials is feasible but sometimes complicated by the presence of calcifications that can appear in very similar fashion. Recently, our group has introduced an automatic method to localize fiducials based on the fiducial's distinct distortion of the local magnetic field, which can be detected on phase images. The accuracy of this method is comparable to CT based The superior soft tissue of MRI over CT greatly facilitates the critical step in the simulation process: the tumor and OAR delineation. Currently, CT is still the master image modality as it provides the information on electron density and bony anatomy. Nowadays, thanks to innovations in MR technology and image processing, this is no longer the case. Accurate electron density maps and reference images can be obtained with MRI in a reliable manner. Thus, from an MRI perspective, the traditionally largest technical obstacles to allow MRI to become the sole imaging modality for treatment simulation has been overcome. It is up to radiotherapy clinics to start using MR-only simulation to improve treatment quality, patient comfort and logistics. localization. Conclusions