Authors (including presenting author) :
Ngai HY(1), Law ML(1), Mak YT(1), Wong WC(1), Ngar YK(1)
Affiliation :
(1) Department of Clinical Oncology, Tuen Mun Hospital
Introduction :
Traditionally, computed tomography (CT) forms the basis for radiotherapy treatment planning as it can provide accurate 3-dimensional information essential for organ delineations and dose calculations. Magnetic resonance imaging (MRI), on the other hand, provides superior soft-tissue contrast as compared to CT and the capability of functional imaging for target delineations. However, MRI does not intrinsically provide the information of electron densities required for dose calculations. MRCAT (Magnetic Resonance for Calculating ATtenuation) from Philips is one of the commercial solutions that can generate synthetic CT images from a single MRI scan, thus enables MR-only treatment planning. MRCAT image generation could fail when the system was unable to classify various tissues, such as due to intra-scan movements or when scanning a different body site which was not included in the model-training process or a phantom which was very different from a human. We have developed a semi-automatic algorithm, a.k.a. TMH MRsCT, for generating synthetic CT from MR, to deal with the cases where MRCAT failed.
Objectives :
We aim to develop a synthetic CT generation software to overcome the limitations imposed on the commercial MRCAT software. The software is going to be applied on cases where MRCAT failed, such as phantom scans for quality assurance. This project focuses on reproducing synthetic CT for human subjects, in particular, the pelvis region.
Methodology :
We discovered a linear function for synthetic CT generation from MRI: H = c*[(1-f)*W - f*F] + s, where H is the Hounsfield Unit (HU) of the output CT image, W is a water-only MR image, F is a fat-only MR image, and c, f and s are parameters. The MR source images, W and F, can be obtained by using DIXON protocol, such as the mDIXON protocol in Philips’ scanners. The same linear function can be used on soft tissues and bony regions with different sets of parameters. The parameters were derived by fitting the function to a set of reference CT images which have been aligned to the MR source images. We used the synthetic CT images from MRCAT as our reference since they are intrinsically perfectly aligned to the source images (TMH MRsCT and MRCAT share the same set of source images). Two sets of parameters are obtained by fitting soft tissues (HU < 90) and bony regions (HU > 90), respectively. Our simple model does not classify different substances, therefore, it relies on manual segmentation of bones to guide the HU assignment. Bone segmentations were done on MIM (MIM Software, USA) and the contours were saved in the format of DICOM-RT structure.
Result & Outcome :
Since the model parameters in TMH MRsCT were derived from MRCAT images, the HU accuracy of MRsCT was evaluated by using MRCAT as the benchmark. Synthetic CT images from MRsCT were subtracted by the corresponding images from MRCAT, and the mean absolute error (MAE) was calculated. Pelvis scans of 5 human test subjects (2 males and 3 females) were used in the evaluation. TMH MRsCT has a MAE of (25.0 ± 5.5) HU and (96.5 ± 34.9) HU over soft tissues and bony regions, respectively. The result showed that the synthetic CT generated by our algorithm was comparable to that by the commercial MRCAT software, and therefore TMH MRsCT could be used as an alternative solution to obtain the density information required for dose plannings.
