@MASTERSTHESIS\{IMM2013-06535,
    author       = "D. Andreasen",
    title        = "Creating a Pseudo{-CT} from {MRI} for {MRI-}only based Radiation Therapy Planning",
    year         = "2013",
    school       = "Technical University of Denmark, {DTU} Compute, {E-}mail: compute@compute.dtu.dk",
    address      = "Matematiktorvet, Building 303{-B,} {DK-}2800 Kgs. Lyngby, Denmark",
    type         = "",
    note         = "{DTU} supervisor: Koen Van Leemput, Ph.D., kvle@dtu.dk, {DTU} Compute",
    url          = "http://www.compute.dtu.dk/English.aspx",
    abstract     = "Background: In the planning process of external radiation therapy, {CT} is used as the main imaging modality. The advantage of using {CT} is that the voxel intensity values are directly related to electron density which is needed for dose calculations. Furthermore, {CT} provides an accurate geometrical representation of bone needed for constructing digitally reconstructed radiographs. In recent years, interest in replacing {CT} with {MRI} in the treatment planning process has emerged. This is due to the fact that {MRI} provides a superior soft tissue contrast; a desirable property that could increase the accuracy of target and risk volume delineation. The challenge in replacing {CT} with {MRI} is that the {MRI} intensity values are not related to electron densities and conventional {MRI} sequences cannot obtain signal from bone.
The purpose of this project was to investigate the use of Gaussian Mixture Regression (GMR) and Random Forest regression (RaFR) for creating a pseudo{-CT} image from {MRI} images. Creating a pseudo{-CT} from {MRI} would eliminate the need for a real {CT} scan and thus facilitate an {MRI-}only work ow in the radiation therapy planning process. The use of {GMR} for pseudo{-CT} creation has previously been reported so the reproducibility of these results was investigated. dUTE and mDixon {MRI} image sets as well as Local Binary Pattern (LBP) feature images were investigated as input to the regression models.
Materials and methods: Head scans of three patients xated for whole brain radiation therapy were acquired on a 1 T open {MRI} scanner with ex coils. dUTE and mDixon image sets were obtained. {CT} head scans were also acquired using a standard protocol. A registration of the {CT} and {MRI} image sets was carried out and {LBP} feature images were derived from the dUTE image sets.
All RaFR and {GMR} models were trained with the dUTE image sets as basic input. Some of the models were trained with an additional mDixon or {LBP} input in order to investigate if these inputs could improve the quality of the predicted pseudo-CT. More specically, the impact of adding the {LBP} input was investigated using RaFR and the impact of adding an mDixon input was investigated using both RaFR and {GMR}. A study of the optimal tree depth for RaFR was also carried out. The quality of the resulting pseudo-CTs was quantied in terms of the prediction deviation, the geometrical accuracy of bone and the dosimetric accuracy.
Results: In the {LBP} input study, the results indicated that using LBPs could improve the quality of the pseudo-CT.
In the mDixon input study, the results suggested that both RaFR and {GMR} models were improved when adding the mDixon input. The improvement was mainly observed in terms of smaller prediction deviations in the bone region of the pseudo-CTs and a greater geometrical accuracy. When comparing RaFR and {GMR,} it was found that using RaFR produced pseudo-CTs with the smallest prediction deviations and greatest geometrical accuracy. In terms of the dosimetric accuracy, the dierence was less clear.
Conclusion: The use of {GMR} and RaFR for creating a pseudo{-CT} image from {MRI} images was investigated. The reproducibility of previously reported results using {GMR} was demonstrated. Furthermore, the impact of adding {LBP} and mDixon inputs to the regression models was demonstrated and showed that an improvement of the pseudo{-CT} could be obtained. The results serves as a motivation for further studies using more data and improved feature images."
}