High-statistics image generation from sparse radiation images by four types of machine-learning models

With the development of nuclear medicine diagnostics and treatment, the demand for image processing techniques has been increasing. Although single photon emission computed tomography (SPECT) and positron emission tomography (PET) are most common, the energy ranges they permit for imaging are limited to either below 300 keV (SPECT) or to 511 keV gamma rays (PET). Recently, Compton cameras have attracted attention, owing to their wide energy range, which stretches from a few hundred keV to several MeV. In this study, we performed Compton camera image processing using four machine-learning (ML) techniques: dictionary learning, U-Net, SRGAN, and AUTOMAP. With these techniques, we tried to reduce the artifacts caused by the sparsity of statistics and improve the signal-to-noise ratio (SNR). Thus, these ML models were trained using image pairs reconstructed from high- and low-statistics images. As a result, we succeeded in generating images similar to the ground truth from low-statistics images. We argue that this technique can be applied not only to Compton camera images but also to other radiation imaging devices. As a future perspective, we mention the possibility of applying our imaging and processing technique to in vivo imaging of alpha-particle internal therapy.