Mice PET/CT Dataset Augmentation Using a 3D-Progressive Growing GAN

In this study, we applied GANs for synthetic 3D mice generation, to be used in PET/CT preclinical deep learning (DL) based research. The lack of available datasets makes it difficult for researchers to apply DL to solve tasks that involve small-animal data, such as emission-based attenuation correction in small-animals’ PET/MR. We applied a 3D-Progressive Growing GAN (3D-PGAN) to generate synthetic CT scans of mice. The 3D-PGAN generated a 64 x 64 x 160 CT scan of a mouse in approximately 0.10 seconds with a Fréchet inception distance (FID), which measures the similarity between two datasets of images, of 103.1 on the coronal plane. For reference, a FID of 0 indicates perfect similarity, and a score of 103.1 is comparable to related works’ results. The generated CT scans resembled real CT scans; however, the generated images were often missing finer structures like a mouse’s ribcage.