Feasibility of Genetic Programming for the Optimization of Tissue-Type-Segmented Maps in the Generation of Synthetic CT in Radiation Therapy Treatment Planning

Abstract

Modern radiation therapy treatment planning has traditionally relied upon computed tomography (CT) in modeling the interaction of megavoltage (MV) photons with the various tissues and organs of the body. CT image data provide both detailed information about individual patient anatomy, as well as a voxel-by-voxel three-dimensional grid of Hounsfield units (HU), which specifies, at all points within the patient, essentially the difference between the attenuation coefficient of the tissue within that voxel from the attenuation coefficient of water, normalized to that of water. From the HU value, the relative electron density can be inferred, and as the relative electron density is the major determinant of the interaction of the tissue with MV photons, the radiation dose distribution can then be calculated. Recently, there has been interest in using magnetic resonance imaging (MR) in lieu of CT, as MR provides superior soft-tissue contrast; however, MR does not provide any electron density information. Various approaches have been essayed to create a synthetic CT (synCT) from the MR data. In the present study, genetic programming was used to construct mappings of MR data to HU data for seven tissue types: bladder, cancellous bone, cortical bone, fat, muscle, prostate, and rectum. These maps were then applied to randomly chosen points in five patient data sets to calculate the synCT HU values, which were then compared with the actual HU values from CT images of those same patients. The method produced mean absolute errors (MAE) of 9.28 HU, 33.24 HU, 75.32 HU, 18.64 HU, 17.12 HU, 11.76 HU, and 18.40 HU for the respective tissue types, and these MAE values are less than those of previous approaches, indicating superior performance. Although the method of the present study does require more manual input, the superior performance is compensatory. Further study is necessary to confirm accuracy on entire MR data sets, and to ensure there is no sample variance effect on the current results.