3D CTGAN: generating 3D heterogeneous tissue textures for virtual phantoms
Medical imaging simulation studies benefit from the use of virtual phantoms since they present a defined ground-truth to evaluate and optimize imaging devices and techniques. Currently, these phantoms are primarily morphologically focused, surface-based representations and do not model the heterogeneous material composition within tissues. We develop a 3D CT conditional generative adversarial network (3D CTGAN) that given a phantom will synthesize heterogeneous CT-based textures within tissues with a focus on primary structures such as bone, muscle, and fat. The model was trained on 378 CT image-segmentation pairs taken from a publicly available dataset and validated using 20 additional pairs. Experimental results showed that the model was able to generate realistic heterogeneous textures for given phantoms. These phantoms were compared with original CT scans and had a mean absolute difference of 38.06 ± 2.86 HU. The structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) were 0.88 ± 0.01 and 29.86 ± 1.51 respectively. These metrics marked an improvement of 29%, 2%, and 7% respectively compared to current methods which define structures with homogeneous composition methods. The 3D CTGAN can allow one to enhance existing computational phantoms with generated textures to improve their level of realism and utility for virtual imaging trials.
