Single-Shot Generative Modelling for 99mTc-Maraciclatide Imaging

A novel radiotracer,
^99m
Tc-maraciclatide, is being investigated as an alternative to MRI and ultrasound for detection and quantification of joint inflammation in rheumatoid arthritis patients. Due to the novelty of this tracer and application, there is a lack of available data to train deep learning models to assist prognostics and diagnostics. In this work we investigate the use of deep generative models for generating realistic synthetic data to augment data for the diagnostic and prognostic model training. We investigate Pix2Pix, PAN and introduce a novel amendment to Pix2Pix using Poisson sampling (PoissonGAN). Our results show PAN and PoissonGAN outperforming the novel SinGAN for our purposes and, in an evaluation by clinicians, the images generated by PAN and PoissonGAN were ranked as more resembling maraciclatide scans than real images 42% of the time.